This PDF file contains the front matter associated with SPIE Proceedings Volume 6512, including the Title Page, Copyright information, Table of Contents, Introduction (if any), and the Conference Committee listing.

3D intra- and inter-subject registration of image volumes is important for tasks that include measurements and quantification of temporal/longitudinal changes, atlas-based segmentation, deriving population averages, or voxel and tensor-based morphometry. A number of methods have been proposed to tackle this problem but few of them have focused on the problem of registering whole body image volumes acquired either from humans or small animals. These image volumes typically contain a large number of articulated structures, which makes registration more difficult than the registration of head images, to which the vast majority of registration algorithms have been applied. To solve this problem, we have previously proposed an approach, which initializes an intensity-based non-rigid registration algorithm with a point based registration technique [1, 2]. In this paper, we introduce new constraints into our non-rigid registration algorithm to prevent the bones from being deformed inaccurately. Results we have obtained show that the new constrained algorithm leads to better registration results than the previous one.

We propose a registration method for the alignment of contrast-enhanced CT liver images. It consists of a fluid-based registration algorithm designed to incorporate a volume-preserving constraint. More specifically our objective is to recover an accurate non-rigid transformation in a perfusion study in presence of contrast-enhanced structures which preserves the incompressibility of liver tissues. This transformation is obtained by integrating a smooth divergence-free vector field derived from the gradient of a statistical similarity measure. This gradient is regularized with a fast recursive low-pass filter and is projected onto the space of divergence-free vector fields using a multigrid solver. Both 2D and 3D versions of the algorithm have been implemented. Simulations and experiments show that our approach improves the registration capture range, enforces the imcompressibility constraint with a good level of accuracy, and is computationally efficient. On perfusion studies, this method prevents the shrinkage of contrast-enhanced regions typically observed with standard fluid methods.

We introduce a new hybrid approach for spline-based elastic image registration using both landmarks and intensity information. As underlying deformation model we use Gaussian elastic body splines (GEBS), which are solutions of the Navier equation of linear elasticity under Gaussian forces. We also incorporate landmark localization uncertainties, which are characterized by weight matrices representing anisotropic errors. To combine landmarks and intensity information, we formulate an energy-minimizing functional that simultaneously minimizes w.r.t. both the landmark and intensity information. The resulting functional can be efficiently minimized using the method of Levenberg/Marquardt. Since the approach is based on a physical deformation model, cross-effects in elastic deformations can be taken into account. We demonstrate the applicability of our scheme based on 3D synthetic images, 2D MR images of the brain, as well as 2D gel electrophoresis images. It turns out that the new scheme achieves more accurate results compared to a pure landmark-based approach.

It is known that along the phase encoding direction the effect of magnetic field inhomogeneity causes significant spatial distortions in fast functional MRI Echo Planar Imaging (EPI). In this work, our previously developed distortion correction technique via a non-rigid registration of EPI to anatomical MRI is improved by adding information from field maps to achieve a more accurate and efficient registration. Local deformation models are used in regions of distortion artifacts instead of using a global non-rigid transformation. The use of local deformations not only enhances the efficiency of the non-rigid registration by reducing the number of deformation model parameters, but also provides constraints to avoid physically incorrect deformations in undistorted regions. The accuracy and reliability of the non-rigid registration technique is improved by using an additional high-resolution gradient echo EPI scan. In-vivo validation is performed by comparing the similarity of the low-resolution EPI to various structural MRI scans before and after applying the computed deformation models. Visual inspection of the images, as well as Mutual Information (MI) and Normalized Cross Correlation (NCC) comparisons, reveal improvements within the sub-voxel range in the moderately distorted areas but not in the signal loss regions.

We propose to use a recently introduced optimisation method in the context of rigid registration of medical images. This optimisation method, introduced by Powell and called NEWUOA, is compared with two other widely used algorithms: Powell's direction set and Nelder-Mead's downhill simplex method. This paper performs a comparative evaluation of the performances of these algorithms to optimise different image similarity measures for different mono- and multi-modal registrations. Images from the BrainWeb project are used as a gold standard for validation purposes. This paper exhibits that the proposed optimisation algorithm is more robust, more accurate and faster than the two other methods.

Medical imaging applications of rigid and non-rigid elastic deformable image registration are undergoing wide scale development. Our approach determines image deformation maps through a hierarchical process, from global to local scales. Vemuri (2000) reported a registration method, based on levelset evolution theory, to morph an image along the motion gradient until it deforms to the reference image. We have applied this level set motion method as basis to iteratively compute the incremental motion fields and then we approximated the field using a higher-level affine and non-rigid motion model. In such a way, we combine sequentially the global affine motion, local affine motion and local non-rigid motion. Our method is fully automated, computationally efficient, and is able to detect large deformations if used together with multi-grid approaches, potentially yielding greater registration accuracy.

A fully automated initialization method is proposed for the 2D/3D registration of 3D CAD models of knee implant components to a single-plane calibrated fluoroscopy. The algorithm matches edge segments, detected in the fluoroscopy image, with pre-computed libraries of expected 2D silhouettes of the implant components. Each library entry represents a different combination of out-of-plane registration transformation parameters. Library matching is performed by computing point-based 2D/2D registrations in between each library entry and each detected edge segment in the fluoroscopy image, resulting in an estimate of the in-plane registration transformation parameters. Point correspondences for registration are established by template matching of the bending patterns on the contours. A matching score for each individual 2D/2D registration is computed by evaluating the transformed library entry in an edge-encoded (characteristic) image, which is derived from the original fluoroscopy image. A matching scores accumulator is introduced to select and suggest one or more initial pose estimates. The proposed method is robust against occlusions and partial segmentations. Validation results are shown on simulated fluoroscopy images. In all cases a library match is found for each implant component which is very similar to the shape information in the fluoroscopy. The feasibility of the proposed method is demonstrated by initializing an intensity-based 2D/3D registration method with the automatically obtained estimation of the registration transformation parameters.

Registration of two dimensional to three dimensional orthopaedic medical image data has important applications particularly in the area of image guided surgery and sports medicine. Fluoroscopy to computer tomography (CT) registration is an important case, wherein digitally reconstructed radiographs derived from the CT data are registered to the fluoroscopy data. Traditional registration metrics such as intensity-based mutual information (MI) typically work well but often suffer from gross misregistration errors when the image to be registered contains a partial view of the anatomy visible in the target image. Phase-based MI provides a robust alternative similarity measure which, in addition to possessing the general robustness and noise immunity that MI provides, also employs local phase information in the registration process which makes it less susceptible to the aforementioned errors. In this paper, we propose using the complex wavelet transform for computing image phase information and incorporating that into a phase-based MI measure for image registration. Tests on a CT volume and 6 fluoroscopy images of the knee are presented. The femur and the tibia in the CT volume were individually registered to the fluoroscopy images using intensity-based MI, gradient-based MI and phase-based MI. Errors in the coordinates of fiducials present in the bone structures were used to assess the accuracy of the different registration schemes. Quantitative results demonstrate that the performance of intensity-based MI was the worst. Gradient-based MI performed slightly better, while phase-based MI results were the best consistently producing the lowest errors.

We are investigating three-dimensional (3D) to two-dimensional (2D) registration methods for computed tomography (CT) and dual-energy digital radiography (DR) for the detection of coronary artery calcification. CT is an established tool for the diagnosis of coronary artery diseases (CADs). Dual-energy digital radiography could be a cost-effective alternative for screening coronary artery calcification. In order to utilize CT as the "gold standard" to evaluate the ability of DR images for the detection and localization of calcium, we developed an automatic intensity-based 3D-to-2D registration method for 3D CT volumes and 2D DR images. To generate digital rendering radiographs (DRR) from the CT volumes, we developed three projection methods, i.e. Gaussian-weighted projection, threshold-based projection, and average-based projection. We tested normalized cross correlation (NCC) and normalized mutual information (NMI) as similarity measurement. We used the Downhill Simplex method as the search strategy. Simulated projection images from CT were fused with the corresponding DR images to evaluate the localization of cardiac calcification. The registration method was evaluated by digital phantoms, physical phantoms, and clinical data sets. The results from the digital phantoms show that the success rate is 100% with mean errors of less 0.8 mm and 0.2 degree for both NCC and NMI. The registration accuracy of the physical phantoms is 0.34 ± 0.27 mm. Color overlay and 3D visualization of the clinical data show that the two images are registered well. This is consistent with the improvement of the NMI values from 0.20 ± 0.03 to 0.25 ± 0.03 after registration. The automatic 3D-to-2D registration method is accurate and robust and may provide a useful tool to evaluate the dual-energy DR images for the detection of coronary artery calcification.

In X-ray guided procedures, the surgeon or interventionalist is dependent on his or her knowledge of the patient's specific anatomy and the projection images acquired during the procedure by a rotational X-ray source. Unfortunately, these X-ray projections fail to give information on the patient's anatomy in the dimension along the projection axis. It would be very profitable to provide the surgeon or interventionalist with a 3D insight of the patient's anatomy that is directly linked to the X-ray images acquired during the procedure. In this paper we present a new robust 2D-3D registration method based on the Projection-Slice Theorem. This theorem gives us a relation between the pre-operative 3D data set and the interventional projection images. Registration is performed by minimizing a translation invariant similarity measure that is applied to the Fourier transforms of the images. The method was tested by performing multiple exhaustive searches on phantom data of the Circle of Willis and on a post-mortem human skull. Validation was performed visually by comparing the test projections to the ones that corresponded to the minimal value of the similarity measure. The Projection-Slice Theorem Based method was shown to be very effective and robust, and provides capture ranges up to 62 degrees. Experiments have shown that the method is capable of retrieving similar results when translations are applied to the projection images.

Due to the long imaging times in SPECT, patient motion is inevitable and constitutes a serious problem for any reconstruction algorithm. The measured inconsistent projection data leads to reconstruction artefacts which can significantly affect the diagnostic accuracy of SPECT, if not corrected. Among the most promising attempts for addressing this cause of artefacts, is the so-called data-driven motion correction methodology. To use this approach it is necessary to automatically detect patient motion and to subdivide the acquired data in projection sets accordingly. In this note, we propose three different schemes for automatically detecting patient motion. All methods were tested on 3D academic examples with different rigid motions, motion times, and camera systems. On the whole, every method was tested with approximately 400 to 600 test cases. One of the proposed new methods does show promising results.

This study was undertaken to correct for motion artifacts in dynamic breast F-18-FDG PET/CT images, to improve differential-image quality, and to increase accuracy of time-activity curves. Dynamic PET studies, with subjects prone, and breast suspended freely employed a protocol with 50 frames, each 1-minute long. A 30 s long CT scan was acquired immediately before the first PET frame. F-18-FDG was administered during the first PET time frame. Fiducial skin markers (FSMs) each containing ~0.5 &mgr;Ci of Ge-68 were taped to each breast. In our PET/PET registration method we utilized CT data. For corresponding FSMs visible on the 1^st and n^th frames, the geometrical centroids of FSMs were found and their displacement vectors were estimated and used to deform the finite element method (FEM) mesh of the CT image (registered with 1^st PET frame) to match the consecutive dynamic PET time frames. Each mesh was then deformed to match the 1^st PET frame using known FSM displacement vectors as FEM loads, and the warped PET timeframe volume was created. All PET time frames were thus nonrigidly registered with the first frame. An analogy between orthogonal components of the displacement field and the temperature distribution in steady-state heat transfer in solids is used, via standard heat-conduction FEM software with "conductivity" of surface elements set arbitrarily significantly higher than that of volume elements. Consequently, the surface reaches steady state before the volume. This prevents creation of concentrated FEM loads at the locations of FSMs and reaching incorrect FEM solution. We observe improved similarity between the 1^st and n^th frames. The contrast and the spatial definition of metabolically hyperactive regions are superior in the registered 3D images compared to unregistered 3D images. Additional work is needed to eliminate small image artifacts due to FSMs.

This paper deals with registering 3D PET images in order to monitor lung tumor evolution. Registering directly two PET images, taken at different stages of a cancer therapy, leads to deforming the tumor of the moving image to take the shape of the fixed image, loosing the tumor evolution information. This results in aberrant medical diagnosis. In order to solve this problem, we propose an indirect registration method that consists of processing pairs of CT-PET images. The CT images acquired at each stage are first registered to estimate anatomical transformations. The free-form deformation obtained is then applied to the corresponding PET images. The reconstructed PET images can be compared and used to monitor the tumor. The volume ratio and radiation density are calculated to assess the evolution of the tumor and evaluate the effectiveness of a therapy. The proposed iconic registration method is based on a B-Spline deformable model and mutual information. Two approaches have been used to validate the proposed method. First, we used phantoms to simulate the evolution of a tumor. The second approach consisted of simulating a tumor within real images. Quantitative measures show that our registration method keeps invariant volume and density distribution ratios of the tumor within PET images. This leads to improved tumor localisation and better evaluation of the efficiency of therapies.

Myocardial perfusion Magnetic Resonance (MR) imaging has proven to be a powerful method to assess coronary artery diseases. The current work presents a novel approach to the analysis of registered sequences of myocardial perfusion MR images. A previously reported active appearance model (AAM) based segmentation and registration of the myocardium provided pixel-wise signal intensity curves that were analyzed using the Support Vector Domain Description (SVDD). In contrast to normal SVDD, the entire regularization path was calculated and used to calculate a generalized distance, which is used to discriminate between ischemic and healthy tissue. The results corresponded well to the ischemic segments found by assessment of the three common perfusion parameters; maximum upslope, peak and time-to-peak obtained pixel-wise.

Deformable models have already been successfully applied to the semi-automatic segmentation of organs from medical images. We present an approach which enables the fully automatic segmentation of the heart from multi-slice computed tomography images. Compared to other approaches, we address the complete segmentation chain comprising both model initialization and adaptation. A multi-compartment mesh describing both atria, both ventricles, the myocardium around the left ventricle and the trunks of the great vessels is adapted to an image volume. The adaptation is performed in a coarse-to- fine manner by progressively relaxing constraints on the degrees of freedom of the allowed deformations. First, the mesh is translated to a rough estimate of the heart's center of mass. Then, the mesh is deformed under the action of image forces. We first constrain the space of deformations to parametric transformations, compensating for global misalignment of the model chambers. Finally, a deformable adaptation is performed to account for more local and subtle variations of the patient's anatomy. The whole heart segmentation was quantitatively evaluated on 25 volume images and qualitatively validated on 42 clinical cases. Our approach was found to work fully automatically in 90% of cases with a mean surface- to-surface error clearly below 1.0 mm. Qualitatively, expert reviewers rated the overall segmentation quality as 4.2±0.7 on a 5-point scale.

Segmentation of organs in medical images can be successfully performed with deformable models. Most approaches combine a boundary detection step with some smoothness or shape constraint. An objective function for the model deformation is thus established from two terms: the first one attracts the surface model to the detected boundaries while the second one keeps the surface smooth or close to expected shapes. In this work, we assign locally varying boundary detection functions to all parts of the surface model. These functions combine an edge detector with local image analysis in order to accept or reject possible edge candidates. The goal is to optimize the discrimination between the wanted and misleading boundaries. We present a method to automatically learn from a representative set of 3D training images which features are optimal at each position of the surface model. The basic idea is to simulate the boundary detection for the given 3D images and to select those features that minimize the distance between the detected position and the desired object boundary. The approach is experimentally evaluated for the complex task of full-heart segmentation in CT images. A cyclic cross-evaluation on 25 cardiac CT images shows that the optimized feature training and selection enables robust, fully automatic heart segmentation with a mean error well below 1 mm. Comparing this approach to simpler training schemes that use the same basic formalism to accept or reject edges shows the importance of the discriminative optimization.

An automatic method is presented to segment the internal carotid arteries through the difficult part of the skull base in CT angiography. The method uses the entropy per slice to select a cross sectional plane below the skull base. In this plane 2D circular structures are detected by the Hough transform. The center points are used to initialize a level set which evolves with a prior shape constraint on its topology. In contrast with some related vessel segmentation methods, our approach does not require the acquisition of an additional CT scan for bone masking. Experiments on twenty internal carotids in ten patients show that 19 seed points are correctly identified (95%) and 18 carotids (90%) are successfully segmented without any human interaction.

Bronchoscopic biopsy of the central-chest lymph nodes is vital in the staging of lung cancer. Three-dimensional multi-detector CT (MDCT) images provide vivid anatomical detail for planning bronchoscopy. Unfortunately, many lymph nodes are situated close to the aorta, and an inadvertent needle biopsy could puncture the aorta, causing serious harm. As an eventual aid for more complete planning of lymph-node biopsy, it is important to define the aorta. This paper proposes a method for extracting the aorta from a 3D MDCT chest image. The method has two main phases: (1) Off-line Model Construction, which provides a set of training cases for fitting new images, and (2) On-Line Aorta Construction, which is used for new incoming 3D MDCT images. Off-Line Model Construction is done once using several representative human MDCT images and consists of the following steps: construct a likelihood image, select control points of the medial axis of the aortic arch, and recompute the control points to obtain a constant-interval medial-axis model. On-Line Aorta Construction consists of the following operations: construct a likelihood image, perform global fitting of the precomputed models to the current case's likelihood image to find the best fitting model, perform local fitting to adjust the medial axis to local data variations, and employ a region recovery method to arrive at the complete constructed 3D aorta. The region recovery method consists of two steps: model-based and region-growing steps. This region growing method can recover regions outside the model coverage and non-circular tube structures. In our experiments, we used three models and achieved satisfactory results on twelve of thirteen test cases.

Brain morphometric studies often incorporate comparative asymmetry analyses of left and right hemispheric brain structures. In this work we show evidence that common methods of user guided structural segmentation exhibit strong left-right asymmetric biases and thus fundamentally influence any left-right asymmetry analyses. We studied several structural segmentation methods with varying degree of user interaction from pure manual outlining to nearly fully automatic procedures. The methods were applied to MR images and their corresponding left-right mirrored images from an adult and a pediatric study. Several expert raters performed the segmentations of all structures. The asymmetric segmentation bias is assessed by comparing the left-right volumetric asymmetry in the original and mirrored datasets, as well as by testing each sides volumetric differences to a zero mean standard t-tests. The structural segmentations of caudate, putamen, globus pallidus, amygdala and hippocampus showed a highly significant asymmetric bias using methods with considerable manual outlining or landmark placement. Only the lateral ventricle segmentation revealed no asymmetric bias due to the high degree of automation and a high intensity contrast on its boundary. Our segmentation methods have been adapted in that they are applied to only one of the hemispheres in an image and its left-right mirrored image. Our work suggests that existing studies of hemispheric asymmetry without similar precautions should be interpreted in a new, skeptical light. Evidence of an asymmetric segmentation bias is novel and unknown to the imaging community. This result seems less surprising to the visual perception community and its likely cause is differences in perception of oppositely curved 3D structures.

Alzheimer's disease (AD) is a neurodegenerative disorder that causes cortical atrophy and impaired cognitive functions. The diagnosis is difficult to make and is often made over a longer period of time using a combination of neuropsychological tests, and structural and functional imaging. Due to the impact of early intervention the challenge of distinguishing early AD from normal ageing has received increasing attention. This study uses cortical thickness measurements to characterize the atrophy in nine mild AD patients (mean MMSE-score 23.3 (std: 2.6)) compared to five healthy middle-aged subjects. A fully automated method based on deformable models is used for delineation of the inner and outer boundaries of the cerebral cortex from Magnetic Resonance Images. This allows observer independent high-resolution quantification of the cortical thickness. The cortex analysis facilitates detection of alterations throughout the entire cortical mantle. To perform inter-subject thickness comparison in which the spatial information is retained, a feature-based registration algorithm is developed which uses local cortical curvature, normal vector, and a distance measure. A comparison of the two study groups reveals that the lateral side of the hemispheres shows diffuse thinner areas in the mild AD group but especially the medial side shows a pronounced thinner area which can be explained by early limbic changes in AD. For classification principal component analysis is applied to reduce the high number of thickness measurements (>200,000) into fewer features. All mild AD and healthy middle-aged subjects are classified correctly (sensitivity and specificity 100%).

We describe a new framework for measuring cortical thickness from MR human brain images. This involves the integration of a method of tissue classification with one to estimate thickness in 3D. We have determined an additional boundary detection step to facilitate this. The classification stage utlizes the Expectation Maximisation (EM) algorithm to classify voxels associated with the tissue types that interface with cortical grey matter (GM, WM and CSF). This uses a Gaussian mixture and the EM algorithm to estimate the position and and width of the Gaussians that model the intensity distributions of the GM, WM and CSF tissue classes. The boundary detection stage uses the GM, WM and CSF classifications and finds connected components, fills holes and then applies a geodesic distance transform to determine the GM/WM interface. Finally the thickness of the cortical grey matter is estimated by solving Laplace's equation and determining the streamlines that connect the inner and outer boundaries. The contribution of this work is the adaptation of the classification and thickness measurement steps, neither requiring manual initialisation, and also the validation strategy. The resultant algorithm is fully automatic and avoids the computational expense associated with preserving the cortical surface topology. We have devised a validation strategy that indicates the cortical segmentation of a gold standard brain atlas has a similarity index of 0.91, thickness estimation has subvoxel accuracy evaluated using a synthetic image and precision of the combined segmentation and thickness measurement of 1.54mm using three clinical images.

In this work we compare 3D Gyrification Index and our recently proposed area-independent curvature-based surface measures [26] for the in-vivo quantification of brain surface folding in clinically acquired neonatal MR image data. A meaningful comparison of gyrification across brains of different sizes and their subregions will only be possible through the quantification of folding with measures that are independent of the area of the region of analysis. This work uses a 3D implementation of the classical Gyrification Index, a 2D measure that quantifies folding based on the ratio of the inner and outer contours of the brain and which has been used to study gyral patterns in adults with schizophrenia, among other conditions. The new surface curvature-based measures and the 3D Gyrification Index were calculated on twelve premature infants (age 28-37 weeks) from which surfaces of cerebrospinal fluid/gray matter (CSF/GM) interface and gray matter/white matter (GM/WM) interface were extracted. Experimental results show that our measures better quantify folding on the CSF/GM interface than Gyrification Index, and perform similarly on the GM/WM interface.

In this work we obtain estimates of tissue growth using longitudinal data comprising MR brain images of 25 preterm children scanned at one and two years. The growth estimates are obtained using segmentation and registration based methods. The segmentation approach used an expectation maximisation (EM) method to classify tissue types and the registration approach used tensor based morphometry (TBM) applied to a free form deformation (FFD) model. The two methods show very good agreement indicating that the registration and segmentation approaches can be used interchangeably. The advantage of the registration based method, however, is that it can provide more local estimates of tissue growth. This is the first longitudinal study of growth in early childhood, previous longitudinal studies have focused on later periods during childhood.

Medical research is dominated by animal models, especially rats and mice. Within a species most laboratory subjects exhibit little variation in brain anatomy. This uniformity of features is used to crop regions of interest based upon a known, cropped brain atlas. For any study involving N subjects, image registration or alignment to an atlas is required to construct a composite result. A highly resolved stack of T2 weighted MRI anatomy images of a Sprague-Dawley rat was registered and cropped to a known segmented atlas. This registered MRI volume was used as the reference atlas. A Pulse Coupled Neural Network (PCNN) was used to separate brain tissue from surrounding structures, such as cranium and muscle. Each iteration of the PCNN produces binary images of increasing area as the intensity spectrum is increased. A rapid filtering algorithm is applied that breaks narrow passages connecting larger segmented areas. A Generalized Invariant Hough Transform is applied subsequently to each PCNN segmented area to identify which segmented reference slice it matches. This process is repeated for multiple slices within each subject. Since we have apriori knowledge of the image ordering and fields of view this information provides initial estimates for subsequent registration codes. This process of subject slice extraction to PCNN mask creations and GIHT matching with known atlas locations is fully automatic.

The research presented in this article is aimed at developing an automated imaging system for classification of normal tissues in medical images obtained from Computed Tomography (CT) scans. Texture features based on a bank of Gabor filters are used to classify the following tissues of interests: liver, spleen, kidney, aorta, trabecular bone, lung, muscle, IP fat, and SQ fat. The approach consists of three steps: convolution of the regions of interest with a bank of 32 Gabor filters (4 frequencies and 8 orientations), extraction of two Gabor texture features per filter (mean and standard deviation), and creation of a Classification and Regression Tree-based classifier that automatically identifies the various tissues. The data set used consists of approximately 1000 DIACOM images from normal chest and abdominal CT scans of five patients. The regions of interest were labeled by expert radiologists. Optimal trees were generated using two techniques: 10-fold cross-validation and splitting of the data set into a training and a testing set. In both cases, perfect classification rules were obtained provided enough images were available for training (~65%). All performance measures (sensitivity, specificity, precision, and accuracy) for all regions of interest were at 100%. This significantly improves previous results that used Wavelet, Ridgelet, and Curvelet texture features, yielding accuracy values in the 85%-98% range The Gabor filters' ability to isolate features at different frequencies and orientations allows for a multi-resolution analysis of texture essential when dealing with, at times, very subtle differences in the texture of tissues in CT scans.

Multi-Slice Computed Tomography (MSCT) imaging of the lungs allow for detection and follow-up of very small lesions including solid and ground glass nodules (GGNs). However relatively few computer-based methods have been implemented for GGN segmentation. GGNs can be divided into pure GGNs and mixed GGNs, which contain both nonsolid and solid components (SC). This latter category is especially of interest since some studies indicate a higher likelihood of malignancy in GGNs with SC. Due to their characteristically slow growth rate, GGNs are typically monitored with multiple follow-up scans, making measurement of the volume of both solid and non-solid component especially desirable. We have developed an automated method to estimate the SC percentage within a segmented GGN. First, the SC algorithm uses a novel method to segment out the solid structures, while excluding any vessels passing near or through the nodule. A gradient distribution analysis around solid structures validates the presence or absence of SC. We tested 50 GGNs, split between three groups: 15 GGNs with SC, 15 GGNs with a solid nodule added to simulate SC, and 20 GGNs without SC. With three defined satisfaction levels for the segmentation (A: succeed, B: acceptable, C: failed), the first group resulted in 60% with score A, 40% with score B, 0% with score C. The second group resulted in 66.7% with score A and 33.3% with score B. In testing the first and 3rd groups, the algorithm correctly detected SC in all cases where it was present (sensitivity of 100%) and correctly determined absence of SC in 15 out of 20 cases (specificity 75%).

Useful diagnosis of lung lesions in computed tomography (CT) depends on many factors including the ability of radiologists to detect and correctly interpret the lesions. Computer-aided Diagnosis (CAD) systems can be used to increase the accuracy of radiologists in this task. CAD systems are, however, trained against ground truth and the mechanisms employed by the CAD algorithms may be distinctly different from the visual perception and analysis tasks of the radiologist. In this paper, we present a framework for finding the mappings between human descriptions and characteristics and computed image features. The data in our study were generated from 29 thoracic CT scans collected by the Lung Image Database Consortium (LIDC). Every case was annotated by up to 4 radiologists by marking the contour of nodules and assigning nine semantic terms to each identified nodule; fifty-nine image features were extracted from each segmented nodule. Correlation analysis and stepwise multiple regression were applied to find correlations among semantic characteristics and image features and to generate prediction models for each characteristic based on image features. From our preliminary experimental results, we found high correlations between different semantic terms (margin, texture), and promising mappings from image features to certain semantic terms (texture, lobulation, spiculation, malignancy). While the framework is presented with respect to the interpretation of pulmonary nodules in CT images, it can be easily extended to find mappings for other modalities in other anatomical structures and for other image features.

Lung CAD systems require the ability to classify a variety of pulmonary structures as part of the diagnostic process. The purpose of this work was to develop a methodology for fully automated voxel-by-voxel classification of airways, fissures, nodules, and vessels from chest CT images using a single feature set and classification method. Twenty-nine thin section CT scans were obtained from the Lung Image Database Consortium (LIDC). Multiple radiologists labeled voxels corresponding to the following structures: airways (trachea to 6th generation), major and minor lobar fissures, nodules, and vessels (hilum to peripheral), and normal lung parenchyma. The labeled data was used in conjunction with a supervised machine learning approach (AdaBoost) to train a set of ensemble classifiers. Each ensemble classifier was trained to detect voxels part of a specific structure (either airway, fissure, nodule, vessel, or parenchyma). The feature set consisted of voxel attenuation and a small number of features based on the eigenvalues of the Hessian matrix (used to differentiate structures by shape) computed at multiple smoothing scales to improve the detection of both large and small structures. When each ensemble classifier was composed of 20 weak classifiers, the AUC values for the airway, fissure, nodule, vessel, and parenchyma classifiers were 0.984 ± 0.011, 0.949 ± 0.009, 0.945 ± 0.018, 0.953 ± 0.016, and 0.931± 0.015 respectively.

Quantification of cerebral blood flow and volume using dynamic-susceptibility contrast MRI relies on deconvolution with the arterial input function (AIF) - commonly estimated from signal changes in a major artery. Manual selection of AIF is user-dependent and typical selection in primary arteries leads to errors due to bolus delay and dispersion. An AIF sampled form the primary as well as the peripheral arteries should minimize these errors. We present a fully automated technique for the identification of the AIF by classifying the pixels in the imaging set into unique classes using a Kohonen self organizing map, followed by an iterative refinement of the previous selections. Validation was performed across 31 pediatric patients by comparison with manually identified AIF and a recently published automated AIF technique. Our technique consistently yielded higher bolus peak heights and over 50% increase in the area under the first pass, therefore lowering the values obtained for blood flow and volume. This technique provides a robust and accurate estimation of the arterial input function and can easily be adapted to extract the AIF locally, regionally or globally as suitable to the analysis.

Level set methods have become increasingly popular as a framework for image segmentation. Yet when used as a generic segmentation tool, they suffer from an important drawback: Current formulations do not allow much user interaction. Upon initialization, boundaries propagate to the final segmentation without the user being able to guide or correct the segmentation. In the present work, we address this limitation by proposing a probabilistic framework for image segmentation which integrates input intensity information and user interaction on equal footings. The resulting algorithm determines the most likely segmentation given the input image and the user input. In order to allow a user interaction in real-time during the segmentation, the algorithm is implemented on a graphics card and in a narrow band formulation.

In the current vast image segmentation literature, there is a serious lack of methods that would allow theoretical comparison of the algorithms introduced by using different mathematical methodologies. The main goal of this article is to introduce a general theoretical framework for image segmentation that would allow such comparison. The framework is based on the formal definitions designed to answer the following fundamental questions: What is the relation between an idealized image and its digital representation? What properties a segmentation algorithm must satisfy to be acknowledged as acceptable? What does it mean that a digital image segmentation algorithm truly approximates an idealized segmentation model? We use the formulated framework to analyze the front propagation (FP) level set algorithm of Malladi, Sethian, and Vemuri and compare it with the fuzzy connectedness family of algorithms. In particular, we prove that the FP algorithm is weakly model-equivalent with the absolute fuzzy connectedness algorithm of Udupa and Samarasekera used with gradient based afinity. Experimental evidence of this equivalence is also provided. The presented theoretical framework can be used to analyze any arbitrary segmentation algorithm. This line of investigation is a subject of our forthcoming work.

The choice of 3D shape representation for anatomical structures determines the effectiveness with which segmentation, visualization, deformation, and shape statistics are performed. Medial axis-based shape representations have attracted considerable attention due to their inherent ability to encode information about the natural geometry of parts of the anatomy. In this paper, we propose a novel approach, based on nonlinear manifold learning, to the parameterization of medial sheets and object surfaces based on the results of skeletonization. For each single-sheet figure in an anatomical structure, we skeletonize the figure, and classify its surface points according to whether they lie on the upper or lower surface, based on their relationship to the skeleton points. We then perform nonlinear dimensionality reduction on the skeleton, upper, and lower surface points, to find the intrinsic 2D coordinate system of each. We then center a planar mesh over each of the low-dimensional representations of the points, and map the meshes back to 3D using the mappings obtained by manifold learning. Correspondence between mesh vertices, established in their intrinsic 2D coordinate spaces, is used in order to compute the thickness vectors emanating from the medial sheet. We show results of our algorithm on real brain and musculoskeletal structures extracted from MRI, as well as an artificial multi-sheet example. The main advantages to this method are its relative simplicity and noniterative nature, and its ability to correctly compute nonintersecting thickness vectors for a medial sheet regardless of both the amount of coincident bending and thickness in the object, and of the incidence of local concavities and convexities in the object's surface.

Image segmentation plays an important role in medical image analysis and visualization since it greatly enhances the clinical diagnosis. Although many algorithms have been proposed, it is challenging to achieve an automatic clinical organ segmentation which requires speed and robustness. Automatically segmenting cardiac Magnetic Resonance Imaging (MRI) image is extremely challenging due to the artifacts of cardiac motion and characteristics of MRI. Moreover many of the existing algorithms are specific to a particular view of cardiac MRI images. We proposed a generic view-independent, learning-based method to automatically segment cardiac MRI images, which uses machine learning techniques and the geometric shape information. A main feature of our contribution is the fact that the proposed algorithm can use a training set containing a mix of various views and is able to successfully segment any given views. The proposed method consists of four stages. First, we partition the input image into a number of image regions based on their intensity characteristics. Then, we calculate the pre-selected feature descriptions for each generated region and use a trained classi.er to learn the conditional probabilities for every pixel based on the calculated features. In this paper, we use the Support Vector Machine (SVM) to train our classifier. The learned conditional probabilities of every pixel are then fed into an energy function to segment the input image. We optimize our energy function with graph cuts. Finally, domain knowledge is applied to verify the segmentation. Experimental results show that this method is very efficient and robust with respect to image views, slices and motion phases. The method also has the potential to be imaging modality independent as the proposed algorithm is not specific to a particular imaging modality.

In quantitative medical imaging and therapy planning, the optimal combination of automated and interactively defined information is crucial for image segmentation methods to be both efficient and effective. We propose to combine an efficient hierarchical region merging scheme that collects per-region statistics across hierarchy levels with a trainable classification engine that facilitates automated region labeling based on an arbitrary number of reference segmentations. When applying the classification engine, we propose to use a corridor of non-classified regions resulting in a sparse labeling with extremely low false-classification rate, and to attribute labels to the remaining basins through successive merging with ready-labeled basins. The proposed hierarchical region merging scheme also permits to efficiently include interactively defined labels. We denominate this general approach as Hybrid Hierarchical Interactive Image Segmentation Scheme (HIS²). More specifically, we present an extension of the Interactive Watershed Transform, which we combine with a trainable two-class Support Vector Machine based on Gaussian radial basis functions. Finally, we present a novel asymmetric marker scheme, which provides a powerful means of regionally correcting remaining inaccuracies while preserving full detail of the automatic labeling procedure. We denominate the complete algorithm as Hybrid Watershed Transform (HWT), which we apply to one challenging segmentation problem in clinical imaging, namely efficient bone removal in large computed tomography angiographic data sets. Efficiency and accuracy of the proposed methodology is evaluated on multi-slice images from nine different sites. As a result, its ability to rapidly and automatically generate robust and precise segmentation results in combination with a versatile manual correction mechanism could be proven without requiring specific anatomical or geometrical models.

Statistical shape priors try to faithfully represent the full range of biological variations in anatomical structures. These priors are now widely used to restrict shapes; obtained in applications like segmentation and registration; to a subspace of plausible shapes. Principle component analysis (PCA) is commonly used to represent modes of shape variations in a training set. In an attempt to face some of the limitations in the PCA-based shape model, this paper describes a new multi-scale shape prior using independent component analysis (ICA) and adaptive wavelet decomposition. Within a best basis selection framework, the proposed method benefits from the multi-scale nature of wavelet packets, and the capability of ICA to capture higher order statistics in wavelet subspaces. The proposed approach is evaluated using contours from digital x-ray images of five vertebrae of human spine. We demonstrate the ability of the proposed shape prior to capture both local and global shape variations, even with limited number of training samples. Our results also show the performance gains of the ICA-based analysis for the wavelet sub-spaces, as compared to PCA-based analysis approach.

Osteoarthritis (OA) is a degenerative joint disease characterized by articular cartilage degradation. A central problem in clinical trials is quantification of progression and early detection of the disease. The accepted standard for evaluating OA progression is to measure the joint space width from radiographs however; there the cartilage is not visible. Recently cartilage volume and thickness measures from MRI are becoming popular, but these measures don't account for the biochemical changes undergoing in the cartilage before cartilage loss even occurs and therefore are not optimal for early detection of OA. As a first step, we quantify cartilage homogeneity (computed as the entropy of the MR intensities) from 114 automatically segmented medial compartments of tibial cartilage sheets from Turbo 3D T₁ sequences, from subjects with no, mild or severe OA symptoms. We show that homogeneity is a more sensitive technique than volume quantification for detecting early OA and for separating healthy individuals from diseased. During OA certain areas of the cartilage are affected more and it is believed that these are the load-bearing regions located at the center of the cartilage. Based on the homogeneity framework we present an automatic technique that partitions the region on the cartilage that contributes to maximum homogeneity discrimination. These regions however, are more towards the noncentral regions of the cartilage. Our observation will provide valuable clues to OA research and may lead to improving treatment efficacy.

The segmentation of the bones from MR images is useful for performing subsequent segmentation and quantitative measurements of cartilage tissue. In this paper, we present a shape based segmentation scheme for the bones that uses texture features derived from the phase and intensity information in the complex MR image. The phase can provide additional information about the tissue interfaces, but due to the phase unwrapping problem, this information is usually discarded. By using a Gabor filter bank on the complex MR image, texture features (including phase) can be extracted without requiring phase unwrapping. These texture features are then analyzed using a support vector machine classifier to obtain probability tissue matches. The segmentation of the bone is fully automatic and performed using a 3D active shape model based approach driven using gradient and texture information. The 3D active shape model is automatically initialized using a robust affine registration. The approach is validated using a database of 18 FLASH MR images that are manually segmented, with an average segmentation overlap (Dice similarity coefficient) of 0.92 compared to 0.9 obtained using the classifier only.

Optic fundus assessment is widely used for diagnosing vascular and non-vascular pathology. Inspection of the retinal vasculature may reveal hypertension, diabetes, arteriosclerosis, cardiovascular disease and stroke. Due to various imaging conditions retinal images may be degraded. Consequently, the enhancement of such images and vessels in them is an important task with direct clinical applications. We propose a novel technique for vessel enhancement in retinal images that is capable of enhancing vessel junctions in addition to linear vessel segments. This is an extension of vessel filters we have previously developed for vessel enhancement in thoracic CT scans. The proposed approach is based on probabilistic models which can discern vessels and junctions. Evaluation shows the proposed filter is better than several known techniques and is comparable to the state of the art when evaluated on a standard dataset. A ridge-based vessel tracking process is applied on the enhanced image to demonstrate the effectiveness of the enhancement filter.

Commercially-available optical coherence tomography (OCT) systems (e.g., Stratus OCT-3) only segment and provide thickness measurements for the total retina on scans of the macula. Since each intraretinal layer may be affected differently by disease, it is desirable to quantify the properties of each layer separately. Thus, we have developed an automated segmentation approach for the separation of the retina on (anisotropic) 3-D macular OCT scans into five layers. Each macular series consisted of six linear radial scans centered at the fovea. Repeated series (up to six, when available) were acquired for each eye and were first registered and averaged together, resulting in a composite image for each angular location. The six surfaces defining the five layers were then found on each 3-D composite image series by transforming the segmentation task into that of finding a minimum-cost closed set in a geometric graph constructed from edge/regional information and a priori-determined surface smoothness and interaction constraints. The method was applied to the macular OCT scans of 12 patients with unilateral anterior ischemic optic neuropathy (corresponding to 24 3-D composite image series). The boundaries were independently defined by two human experts on one raw scan of each eye. Using the average of the experts' tracings as a reference standard resulted in an overall mean unsigned border positioning error of 6.7 ± 4.0 &mgr;m, with five of the six surfaces showing significantly lower mean errors than those computed between the two observers (p < 0.05, pixel size of 50 × 2 &mgr;m).

Early detection of glaucoma is essential to minimizing the risk of visual loss. It has been shown that a good predictor of glaucoma is the cup-to-disc ratio of the optic nerve head. This paper presents an automated method to segment the optic disc. Our approach utilizes pixel feature selection to train a feature set to recognize the region of the disc. Soft pixel classification is used to generate a probability map of the disc. A new cost function is developed for maximizing the probability of the region within the disc. The segmentation of the image is done using a novel graph search algorithm capable of detecting the border maximizing the probability of the disc. The combination of graph search and pixel classification enables us to incorporate large feature sets into the cost function design, which is critical for segmentation of the optic disc. Our results are validated against a reference standard of 82 datasets and compared to the manual segmentations of 3 glaucoma fellows.

Delineating critical structures for radiotherapy of the brain is required for advanced radiotherapy technologies to determine if the dose from the proposed treatment will impair the functionality of the structures. Employing an automatic segmentation computer module in the radiation oncology treatment planning process has the potential to significantly increase the efficiency, cost-effectiveness, and, ultimately, clinical outcome of patients undergoing radiation therapy. In earlier work, we have shown that atlas-based segmentation of large structures such as the brainstem or the cerebellum was an achievable objective. We have also shown that smaller structures such as the optic nerves or optic chiasm were more difficult to segment automatically. In this work, we present an extension to this approach in which atlas-based segmentation is followed by a series of additional steps. We show that this new approach substantially improves our previous results. We also show that we can segment CT images alone when we previously relied on a combination of MR and CT images.

In this paper we describe a new general tumor segmentation approach, which combines energy minimization methods with radial basis function surface modelling techniques. A tumor is mathematically described by a superposition of radial basis functions. In order to find the optimal segmentation we minimize a certain energy functional. Similar to snake segmentation our energy functional is a weighted sum of an internal and an external energy. The internal energy is the bending energy of the surface and can be computed from the radial basis function coefficients directly. Unlike to snake segmentation we do not have to derive and solve Euler-Lagrange equations. We can solve the minimization problem by standard optimization techniques. Our approach is not restricted to one single imaging modality and it can be applied to 2D, 3D or even 4D data. In addition, our segmentation method makes several simple and intuitive user interactions possible. For instance, we can enforce interpolation of certain user defined points. We validate our new method with lung nodules on CT data. A validation on clinical data is carried out with the 91 publicly available CT lung images provided by the lung image database consortium (LIDC). The LIDC also provides ground truth lists by 4 different radiologists. We discuss the inter-observer variability of the 4 radiologists and compare their segmentations with the segmentation results of the presented algorithm.

Bayesian classification methods have been extensively used in a variety of image processing applications, including medical image analysis. The basic procedure is to combine data-driven knowledge in the likelihood terms with clinical knowledge in the prior terms to classify an image into a pre-determined number of classes. In many applications, it is difficult to construct meaningful priors and, hence, homogeneous priors are assumed. In this paper, we show how expectation-maximization weights and neighboring posterior probabilities may be combined to make intuitive use of the Bayesian priors. Drawing upon insights from computer vision tracking algorithms, we cast the problem in a tissue tracking framework. We show results of our algorithm on the classification of gray and white matter along with surrounding cerebral spinal fluid in brain MRI scans. We show results of our algorithm on 20 brain MRI datasets along with validation against expert manual segmentations.

A novel method for classification and quantification of vertebral fractures from X-ray images is presented. Using pairwise conditional shape models trained on a set of healthy spines, the most likely unfractured shape is estimated for each of the vertebrae in the image. The difference between the true shape and the reconstructed normal shape is an indicator for the shape abnormality. A statistical classification scheme with the two shapes as features is applied to detect, classify, and grade various types of deformities. In contrast with the current (semi-)quantitative grading strategies this method takes the full shape into account, it uses a patient-specific reference by combining population-based information on biological variation in vertebra shape and vertebra interrelations, and it provides a continuous measure of deformity. Good agreement with manual classification and grading is demonstrated on 204 lateral spine radiographs with in total 89 fractures.

A main focus of statistical shape analysis is the description of variability of a population of geometric objects. In this paper, we present work towards modeling the shape and pose variability of sets of multiple objects. Principal geodesic analysis (PGA) is the extension of the standard technique of principal component analysis (PCA) into the nonlinear Riemannian symmetric space of pose and our medial m-rep shape description, a space in which use of PCA would be incorrect. In this paper, we discuss the decoupling of pose and shape in multi-object sets using different normalization settings. Further, we introduce methods of describing the statistics of object pose and object shape, both separately and simultaneously using a novel extension of PGA. We demonstrate our methods in an application to a longitudinal pediatric autism study with object sets of 10 subcortical structures in a population of 47 subjects. The results show that global scale accounts for most of the major mode of variation across time. Furthermore, the PGA components and the corresponding distribution of different subject groups vary significantly depending on the choice of normalization, which illustrates the importance of global and local pose alignment in multi-object shape analysis. Finally, we present results of using distance weighted discrimination analysis (DWD) in an attempt to use pose and shape features to separate subjects according to diagnosis, as well as visualize discriminating differences.

Recent advancements in digital medical imaging have opened avenues for quantitative analyses of different volumetric and morphometric indices in response to a disease or a treatment. However, a major challenge in performing such an analysis is the lack of a technology of building a mean anatomic space (MAS) that allows mapping data of a given subject onto MAS. This approach leads to a tool for point-by-point regional analysis and comparison of quantitative indices for data coming from a longitudinal or transverse study. Toward this goal, we develop a new computation technique, called Active Index Model (AIM), which is a unique tool to solve the stated problem. AIM consists of three building blocks - (1) development of MAS for a particular anatomic site, (2) mapping a specific data onto MAS, (3) regional statistical analysis of data from different populations assessing regional response to a disease or treatment progression. The AIM presented here is built at the training phase from two known populations (e.g., normal and diseased) which will be immediately ready for diagnostic purpose in a subject whose clinical status is unknown. AIM will be useful for both cross sectional and longitudinal studies and for early diagnostic. This technique will be a vital tool for understanding regional response of a disease or treatment at various stages of its progression. This method has been applied for analyzing regional trabecular bone structural distribution in rabbit femur via micro-CT imaging and to localize the affected myocardial regions from cardiac MR data.

Crouzon syndrome is characterised by premature fusion of cranial sutures and synchondroses leading to craniofacial growth disturbances. The gene causing the syndrome was discovered approximately a decade ago and recently the first mouse model of the syndrome was generated. In this study, a set of Micro CT scans of the heads of wild-type (normal) mice and Crouzon mice were investigated. Statistical deformation models were built to assess the anatomical differences between the groups, as well as the within-group anatomical variation. Following the approach by Rueckert et al. we built an atlas using B-spline-based nonrigid registration and subsequently, the atlas was nonrigidly registered to the cases being modelled. The parameters of these registrations were then used as input to a PCA. Using different sets of registration parameters, different models were constructed to describe (i) the difference between the two groups in anatomical variation and (ii) the within-group variation. These models confirmed many known traits in the wild-type and Crouzon mouse craniofacial anatomy. However, they also showed some new traits.

In this paper we present a novel approach to the problem of fitting a 4D statistical shape model of the myocardium to cardiac MR and CT image sequences. The 4D statistical model has been constructed from 25 cardiac MR image sequences from normal volunteers. The model is controlled by two sets of shape parameters. The first set of shape parameters describes shape changes due to inter-subject variability while the second set of shape parameters describes shape changes due to intra-subject variability, i.e. the cardiac contraction and relaxation. A novel fitting approach is used to estimate the optimal parameters of the cardiac shape model. The fitting of the model is performed simultaneously for the entire image sequences. The method has been tested on 5 cardiac MR image sequences. Furthermore, we have also tested the method using a cardiac CT image sequence. The result demonstrate that the method is not only able to fit the 4D model to cardiac MR image sequences, but also to cardiac image sequences from a different modality (CT).

This paper summarises the work we have been doing on joint projects with GE Healthcare on colorectal and liver cancer, and with Siemens Molecular Imaging on dynamic PET. First, we recall the salient facts about cancer and oncological image analysis. Then we introduce some of the work that we have done on analysing clinical MRI images of colorectal and liver cancer, specifically the detection of lymph nodes and segmentation of the circumferential resection margin. In the second part of the paper, we shift attention to the complementary aspect of molecular image analysis, illustrating our approach with some recent work on: tumour acidosis, tumour hypoxia, and multiply drug resistant tumours.

Numerous studies have investigated the relation between mammographic density and breast cancer risk. These studies indicate that women with dense breasts have a four to six fold risk increase. There is currently no gold standard for automatic assessment of mammographic density. In previous work two different automated methods for measuring the effect of HRT w.r.t. changes in breast density have been presented. One is a percentage density based on an adaptive global threshold, and the other is an intensity invariant measure, which provides structural information orthogonal to intensity-based methods. In this article we investigate the ability to detect density changes induced by HRT for these measures and compare to a radiologist's BI-RADS rating and interactive threshold percentage density. In the experiments, two sets of mammograms of 80 patients from a double blind, placebo controlled HRT experiment are used. The p-values for the statistical significance of the separation of density means, for the HRT group and the placebo group at end of study, are 0.2, 0.1, 0.02 and 0.02 for the automatic threshold, BI-RADS, the stripyness and the interactive threshold respectively.

Voxel-based morphometry (VBM) is an automated and objective image analysis technique for detecting differences in regional concentration or volume of brain tissue composition based on structural magnetic resonance (MR) images. VBM has been used widely to evaluate brain morphometric differences between different populations, but there isn't an evaluation system for its validation until now. In this study, a quantitative and objective evaluation system was established in order to assess VBM performance. We recruited twenty normal volunteers (10 males and 10 females, age range 20-26 years, mean age 22.6 years). Firstly, several focal lesions (hippocampus, frontal lobe, anterior cingulate, back of hippocampus, back of anterior cingulate) were simulated in selected brain regions using real MRI data. Secondly, optimized VBM was performed to detect structural differences between groups. Thirdly, one-way ANOVA and post-hoc test were used to assess the accuracy and sensitivity of VBM analysis. The results revealed that VBM was a good detective tool in majority of brain regions, even in controversial brain region such as hippocampus in VBM study. Generally speaking, much more severity of focal lesion was, better VBM performance was. However size of focal lesion had little effects on VBM analysis.

Intensity Modulated Radiotherapy is a new technique enabling the sculpting of the 3D radiation dose. It enables to modulate the delivery of the dose inside the malignant areas and constrain the radiation plan for protecting important functional areas. It also raises the issues of adequacy and accuracy of the selection and delineation of the target volumes. The delineation in the patient image of the tumor volume is highly time-consuming and requires considerable expertise. In this paper we focus on atlas based automatic segmentation of head and neck patients and compare two non-rigid registration methods: B-Spline and Morphons. To assess the quality of each method, we took a set of four 3D CT patient's images previously segmented by a doctor with the organs at risk. After a preliminary affine registration, both non-rigid registration algorithms were applied to match the patient and atlas images. Each deformation field, resulted from the non-rigid deformation, was applied on the masks corresponding to segmented regions in the atlas. The atlas based segmentation masks were compared to manual segmentations performed by an expert. We conclude that Morphons has performed better for matching all structures being considered, improving in average 11% the segmentation.

We propose a method for the automated computation of the mid-sagittal plane of the brain in diffusion tensor MR images. We estimate this plane as the one that best superposes the two hemispheres of the brain by reflection symmetry. This is done via the automated minimisation of a correlation-type global criterion over the tensor image. The minimisation is performed using the NEWUOA algorithm in a multiresolution framework. We validate our algorithm on synthetic diffusion tensor MR images. We quantitatively compare this computed plane with similar planes obtained from scalar diffusion images (such as FA and ADC maps) and from the B0 image (that is, without diffusion sensitisation). Finally, we show some results on real diffusion tensor MR images.

Diffusion Tensor Imaging (DTI) is currently a widespread technique to infer white matter architecture in the human brain. An important application of DTI is to understand the anatomical coupling between functional cortical regions of the brain. To solve this problem, anisotropy maps are insufficient and fiber tracking methods are used to obtain the main fibers. While the diffusion tensor (DT) is important to obtain anisotropy maps and apparent diffusivity of the underlying tissue, fiber tractography using the full DT may result in diffusive tracking that leaks into unexpected regions. Sharpening is thus of utmost importance to obtain complete and accurate tracts. In the tracking literature, only heuristic methods have been proposed to deal with this problem. We propose a new tensor sharpening transform. Analogously to the general issue with the diffusion and fiberOrientation Distribution Function (ODF) encountered when working with High Angular Resolution Diffusion Imaging (HARDI), we show how to transform the diffusion tensors into so-called fiber tensors. We demonstrate that this tensor transform is a natural pre-processing task when one is interested in fiber tracking. It also leads to a dramatic improvement of the tractography results obtained by front propagation techniques on the full diffusion tensor. We compare and validate sharpening and tracking results on synthetic data and on known fiber bundles in the human brain.

An important problem in medical image analysis is the segmentation of anatomical regions of interest. Once regions of interest are segmented, one can extract shape, appearance, and structural features that can be analyzed for disease diagnosis or treatment evaluation. Diffusion tensor magnetic resonance imaging (DT-MRI) is a relatively new medical imaging modality that captures unique water diffusion properties and fiber orientation information of the imaged tissues. In this paper, we extend the interactive multidimensional graph cuts segmentation technique to operate on DT-MRI data by utilizing latest advances in tensor calculus and diffusion tensor dissimilarity metrics. The user interactively selects certain tensors as object ("obj") or background ("bkg") to provide hard constraints for the segmentation. Additional soft constraints incorporate information about both regional tissue diffusion as well as boundaries between tissues of different diffusion properties. Graph cuts are used to find globally optimal segmentation of the underlying 3D DT-MR image among all segmentations satisfying the constraints. We develop a graph structure from the underlying DT-MR image with the tensor voxels corresponding to the graph vertices and with graph edge weights computed using either Log-Euclidean or the J-divergence tensor dissimilarity metric. The topology of our segmentation is unrestricted and both obj and bkg segments may consist of several isolated parts. We test our method on synthetic DT data and apply it to real 2D and 3D MRI, providing segmentations of the corpus callosum in the brain and the ventricles of the heart.

MRI at high magnetic fields (> 3.0 T) is complicated by strong inhomogeneous radio-frequency fields, sometimes termed the "bias field". These lead to nonuniformity of image intensity, greatly complicating further analysis such as registration and segmentation. Existing methods for bias field correction are effective for 1.5 T or 3.0 T MRI, but are not completely satisfactory for higher field data. This paper develops an effective bias field correction for high field MRI based on the assumption that the nonuniformity is smoothly varying in space. Also, nonuniformity is quantified and unmixed using high order neighborhood statistics of intensity cooccurrences. They are computed within spherical windows of limited size over the entire image. The restoration is iterative and makes use of a novel stable stopping criterion that depends on the scaled entropy of the cooccurrence statistics, which is a non monotonic function of the iterations; the Shannon entropy of the cooccurrence statistics normalized to the effective dynamic range of the image. The algorithm restores whole head data, is robust to intense nonuniformities present in high field acquisitions, and is robust to variations in anatomy. This algorithm significantly improves bias field correction in comparison to N3 on phantom 1.5 T head data and high field 4 T human head data.

Echo planar images (EPI) suffer from geometric distortions caused by static-field inhomogeneity. Correction techniques have been suggested based on field maps obtained before or after the EPI acquisition. However, when a relatively long time series of images is required, as in fMRI studies, the inhomogeneity varies from image to image because of gross motion and physiological activity such as respiration and cardiac motion. It is not ideal to approximate the varying maps of field inhomogeneity by means of one map. To overcome this limitation, multiple field maps are desirable for correcting the distortions that are dynamically changing. Some groups have explored the possibility of acquiring multiple field maps, but either the increased scan time is not affordable for most fMRI studies or the field map acquisition is embedded in EPI pulse sequence, which produces a map of insufficient resolution to support a complete distortion correction. In this paper, we propose a dynamic field mapping technique that uses a single reference image and a single corresponding acquired field map and the phase information extracted from the complex image data of each EPI image in the time series. From this information, a separate field map is then derived individually for each EPI image. The derived field maps are then used for distortion correction. This approach, which is particularly suitable for fMRI studies, can correct for image distortion that varies dynamically without sacrificing temporal resolution. We validate this technique using simulated data, and the experimental results show improved performance in comparison to correction using a single field map.

The mobility of lung tumors during the respiratory cycle is a source of error in radiotherapy treatment planning. Spatiotemporal CT data sets can be used for studying the motion of lung tumors and inner organs during the breathing cycle. We present methods for the analysis of respiratory motion using 4D CT data in high temporal resolution. An optical flow based reconstruction method was used to generate artifact-reduced 4D CT data sets of lung cancer patients. The reconstructed 4D CT data sets were segmented and the respiratory motion of tumors and inner organs was analyzed. A non-linear registration algorithm is used to calculate the velocity field between consecutive time frames of the 4D data. The resulting velocity field is used to analyze trajectories of landmarks and surface points. By this technique, the maximum displacement of any surface point is calculated, and regions with large respiratory motion are marked. To describe the tumor mobility the motion of the lung tumor center in three orthogonal directions is displayed. Estimated 3D appearance probabilities visualize the movement of the tumor during the respiratory cycle in one static image. Furthermore, correlations between trajectories of the skin surface and the trajectory of the tumor center are determined and skin regions are identified which are suitable for prediction of the internal tumor motion. The results of the motion analysis indicate that the described methods are suitable to gain insight into the spatiotemporal behavior of anatomical and pathological structures during the respiratory cycle.

Complex challenges of optical imaging in diagnostics and surgical treatment require accurate image registration/stabilization methods that remove only unwanted motions. An SIAROI algorithm is proposed for real-time subpixel registration sequences of intraoperatively acquired infrared (thermal) brain images. SIAROI algorithm is based upon automatic, localized Subpixel Image Autocorrelation and a user-selected Region of Interest (ROI). Human expertise about unwanted motions is added through a user-outlined ROI, using a low-accuracy free-hand paintbrush. SIAROI includes: (a) propagating the user-outlined ROI by selecting pixels in the second image of the sequence, using the same ROI; (b) producing SROI (sub-pixel ROI) by converting each pixel to k=NxN subpixels; (c) producing new SROI in the second image by shifting SROI within plus or minus 6k subpixels; (d) finding an optimal autocorrelation shift (x,y) within 12N that minimizes the Standard Deviation of Differences of Pixel Intensities (SDDPI) between corresponding ROI pixels in both images, (e) shifting the second image by (x,y), repeating (a)-(e) for successive images (t,t1). In experiments, a user quickly outlined non-deformable ROI (such as bone) in the first image of a sequence. Alignment of 100 brain images (25600x25600 pixel search, after every pixel was converted to 100 sub-pixels), took ~3 minutes, which is 200 times faster (with a 0.1=ROI/image ratio) than global auto-correlation. SIAROI improved frame alignment by a factor of two, relative to a Global Auto-correlation and Tie-points-based registration methods, as measured by reductions in the SDDPI.

A system for the semi-automatic reconstruction of highly fragmented bone fractures, developed to aid in treatment planning, is presented. The system aligns bone fragment surfaces derived from segmentation of volumetric CT scan data. Each fragment surface is partitioned into intact- and fracture-surfaces, corresponding more or less to cortical and cancellous bone, respectively. A user then interactively selects fracture-surface patches in pairs that coarsely correspond. A final optimization step is performed automatically to solve the N-body rigid alignment problem. The work represents the first example of a 3D bone fracture reconstruction system and addresses two new problems unique to the reconstruction of fractured bones: (1) non-stationary noise inherent in surfaces generated from a difficult segmentation problem and (2) the possibility that a single fracture surface on a fragment may correspond to many other fragments.

Quantitative measurement of vertebral rotation is important in surgical planning, analysis of surgical results, and monitoring of the progression of spinal deformities. However, many established and newly developed techniques for measuring axial vertebral rotation do not exploit three-dimensional (3D) information, which may result in virtual axial rotation because of the sagittal and coronal rotation of vertebrae. We propose a novel automatic approach to the measurement of the location and rotation of vertebrae in 3D without prior volume reformation, identification of appropriate cross-sections or aid by statistical models. The vertebra under investigation is encompassed by a mask in the form of an elliptical cylinder in 3D, defined by its center of rotation and the rotation angles. We exploit the natural symmetry of the vertebral body, vertebral column and vertebral canal by dividing the vertebral mask by its mid-axial, mid-sagittal and mid-coronal plane, so that the obtained volume pairs contain symmetrical parts of the observed anatomy. Mirror volume pairs are then simultaneously registered to each other by robust rigid auto-registration, using the weighted sum of absolute differences between the intensities of the corresponding volume pairs as the similarity measure. The method was evaluated on 50 lumbar vertebrae from normal and scoliotic computed tomography (CT) spinal scans, showing relatively large capture ranges and distinctive maxima at the correct locations and rotation angles. The proposed method may aid the measurement of the dimensions of vertebral pedicles, foraminae and canal, and may be a valuable tool for clinical evaluation of the spinal deformities in 3D.

A bottleneck for high-throughput screening of live cells is the automated analysis of the generated image data. An important application in this context is the evaluation of the duration of cell cycle phases from confocal time-lapse microscopy image sequences, which typically involves a tracking step. The tracking step is an important part since it relates segmented cells from one time frame to the next. However, a main problem is that often the movement of single cells is superimposed with a global movement. The reason for the global movement lies in the high-throughput acquisition of the images and the repositioning of the microscope. If a tracking algorithm is applied to these images then only a superposition of the microscope movement and the cell movement is determined but not the real movement of the cells. In addition, since the displacements are generally larger, it is more difficult to determine the correspondences between cells. We have developed a phase-correlation based approach to compensate for the global movement of the microscope by registering each image of a sequence to a reference coordinate system. Our approach uses a windowing function in the spatial domain of the cross-power spectrum. This allows to determine the global movement by direct evaluation of the phase gradient, avoiding phase unwrapping. We present experimental results of applying our approach to synthetic and real image sequences. It turns out that the global movement can well be compensated and thus successfully decouples the global movement from the individual movement of the cells.

In this paper we present novel methods for registration and subsequent evaluation of whole mount prostate histological sections to corresponding 4 Tesla ex vivo magnetic resonance imaging (MRI) slices to complement our existing computer-aided diagnosis (CAD) system for detection of prostatic adenocarcinoma from high resolution MRI. The CAD system is trained using voxels labeled as cancer on MRI by experts who visually aligned histology with MRI. To address voxel labeling errors on account of manual alignment and delineation, we have developed a registration method called combined feature ensemble mutual information (COFEMI) to automatically map spatial extent of prostate cancer from histology onto corresponding MRI for prostatectomy specimens. Our method improves over intensity-based similarity metrics (mutual information) by incorporating unique information from feature spaces that are relatively robust to intensity artifacts and which accentuate the structural details in the target and template images to be registered. Our registration algorithm accounts for linear gland deformations in the histological sections resulting from gland fixing and serial sectioning. Following automatic registration of MRI and histology, cancer extent from histological sections are mapped to the corresponding registered MRI slices. The manually delineated cancer areas on MRI obtained via manual alignment of histological sections and MRI are compared with corresponding cancer extent obtained via COFEMI by a novel registration evaluation technique based on use of non-linear dimensionality reduction (locally linear embedding (LLE)). The cancer map on MRI determined by COFEMI was found to be significantly more accurate compared to the manually determined cancer mask. The performance of COFEMI was also found to be superior compared to image intensity-based mutual information registration.

The study was conducted to develop a simple model for more robust lung registration of volumetric CT data, which is essential for various clinical lung analysis applications, including the lung nodule matching in follow up CT studies, semi-quantitative assessment of lung perfusion, and etc. The purpose of this study is to find the most effective reference point and geometric model based on the lung motion analysis from the CT data sets obtained in full inspiration (In.) and expiration (Ex.). Ten pairs of CT data sets in normal subjects obtained in full In. and Ex. were used in this study. Two radiologists were requested to draw 20 points representing the subpleural point of the central axis in each segment. The apex, hilar point, and center of inertia (COI) of each unilateral lung were proposed as the reference point. To evaluate optimal expansion point, non-linear optimization without constraints was employed. The objective function is sum of distances from the line, consist of the corresponding points between In. and Ex. to the optimal point x. By using the nonlinear optimization, the optimal points was evaluated and compared between reference points. The average distance between the optimal point and each line segment revealed that the balloon model was more suitable to explain the lung expansion model. This lung motion analysis based on vector analysis and non-linear optimization shows that balloon model centered on the center of inertia of lung is most effective geometric model to explain lung expansion by breathing.

The purpose of this study was to develop a computerized method for measurement of displacement vectors of target position on electronic portal imaging device (EPID) cine images in a treatment without implanted markers. Our proposed method was based on a template matching technique with cross-correlation coefficient between a reference portal (RP) image and each consecutive portal (CP) image acquired by the EPID. EPID images with 512×384 pixels (pixel size:0.56 mm) were acquired in a cine mode at a sampling rate of 0.5 frame/sec by using an energy of 4, 6, or 10MV on linear accelerators. The displacement vector of the target on each cine image was determined from the position in which took the maximum cross-correlation value between the RP image and each CP image. We applied our method to EPID cine images of a lung phantom with a tumor model simulating respiratory motion, and 5 cases with a non-small cell lung cancer and one case of metastasis. For validation of our proposed method, displacement vectors of a target position calculated by our method were compared with those determined manually by two radiation oncologists. As a result, for lung phantom images, target displacements by our method correlated well with those by the oncologists (r=0.972 - 0.994). Correlation values for 4 cases ranged from 0.854 to 0.991, but the values for the other two cases were 0.609 and 0.644. This preliminary result suggested that our method may be useful for monitoring of displacement vectors of target positions without implanted markers in stereotactic radiotherapy.

High-throughput live-cell imaging is one of the important tools for the investigation of cellular structure and functions in modern experimental biology. Automatic processing of time series of microscopic images is hampered by a number of technical and natural factors such as permanent movements of cells in the optical field, alteration of optical cell appearance and high level of noise. Detection and compensation of global motion of groups of cells or relocation of a single cell within a dynamical multi-cell environment is the first indispensable step in the image analysis chain. This article presents an approach for detection of global image motion and single cell tracking in time series of confocal laser scanning microscopy images using an extended Fourier-phase correlation technique, which allows for analysis of non-uniform multi-body motion in partially-similar images. Our experimental results have shown that the developed approach is capable to perform cell tracking and registration in dynamical and noisy scenes, and provides a robust tool for fully-automatic registration of time-series of microscopic images.

Image registration is an important task in most medical imaging applications. Numerous algorithms have been proposed and some are widely used. However, due to the vast amount of data collected by eg. a computed tomography (CT) scanner, most registration algorithms are very slow and memory consuming. This is a huge problem especially in atlas building, where potentially hundreds of registrations are performed. This paper describes an approach for accelerated image registration. A grid-based warp function proposed by Cootes and Twining, parameterized by the displacement of the grid-nodes, is used. Using a coarse-to-fine approach, the composition of small diffeomorphic warps, results in a final diffeomorphic warp. Normally the registration is done using a standard gradient-based optimizer, but to obtain a fast algorithm the optimization is formulated in the inverse compositional framework proposed by Baker and Matthews. By switching the roles of the target and the input volume, the Jacobian and the Hessian can be pre-calculated resulting in a very efficient optimization algorithm. By exploiting the local nature of the grid-based warp, the storage requirements of the Jacobian and the Hessian can be minimized. Furthermore, it is shown that additional constraints on the registration, such as the location of markers, are easily embedded in the optimization. The method is applied on volumes built from CT-scans of pig-carcasses, and results show a two-fold increase in speed using the inverse compositional approach versus the traditional gradient-based method.

Image registration is now a well understood problem and several techniques using a combination of cost functions, transformation models and optimizers have been reported in medical imaging literature. Parametric methods often rely on the efficient placement of control points in the images, that is, depending on the location and scale at which images are mismatched. Poor choice of parameterization results in deformations not being modeled accurately or over parameterization, where control points may lie in homogeneous regions with low sensitivity to cost. This lowers computational efficiency due to the high complexity of the search space and might also provide transformations that are not physically meaningful, and possibly folded. Adaptive methods that parameterize based on mismatch in images have been proposed. In such methods, the cost measure must be normalized, heuristics such as how many points to pick, resolution of the grids, choosing gradient thresholds and when to refine scale would have to be ascertained in addition to the limitation of working only at a few discrete scales. In this paper we identify mismatch by searching the entire image and a wide range of smooth spatial scales. The mismatch vector, containing location and scale of mismatch is computed from peaks in the local joint entropy. Results show that this method can be used to quickly and effectively locate mismatched regions in images where control points can be placed in preference to other regions speeding up registration.

This paper proposes an extension to the standard iterative closest point method (ICP). In contrast to ICP, our approach (ICP-M) uses the Mahalanobis distance to align a set of shapes thus assigning an anisotropic independent Gaussian noise to each point in the reference shape. The paper introduces the notion of a mahalanobis distance map upon a point set with associated covariance matrices which in addition to providing correlation weighted distance implicitly provides a method for assigning correspondence during alignment. This distance map provides an easy formulation of the ICP problem that permits a fast optimization. Initially, the covariance matrices are set to the identity matrix, and all shapes are aligned to a randomly selected shape (equivalent to standard ICP). From this point the algorithm iterates between the steps: (a) obtain mean shape and new estimates of the covariance matrices from the aligned shapes, (b) align shapes to the mean shape. Three different methods for estimating the mean shape with associated covariance matrices are explored in the paper. The proposed methods are validated experimentally on two separate datasets (IMM face dataset and femur-bones). The superiority of ICP-M compared with ICP in recovering the underlying correspondences in the face dataset is demonstrated.

Multi-spectral images of human tissue taken in-vivo often contain image alignment problems as patients have difficulty in retaining their posture during the acquisition time of 20 seconds. Previously, it has been attempted to correct motion errors with image registration software developed for MR or CT data but these algorithms have been proven to be too slow and erroneous for practical use with multi-spectral images. A new software package has been developed which allows the user to play a decisive role in the registration process as the user can monitor the progress of the registration continuously and force it in the right direction when it starts to fail. The software efficiently exploits videocard hardware to gain speed and to provide a perfect subvoxel correspondence between registration field and display. An 8 bit graphic card was used to efficiently register and resample 12 bit images using the hardware interpolation modes present on the graphic card. To show the feasibility of this new registration process, the software was applied in clinical practice evaluating the dosimetry for psoriasis and KTP laser treatment. The microscopic differences between images of normal skin and skin exposed to UV light proved that an affine registration step including zooming and slanting is critical for a subsequent elastic match to have success. The combination of user interactive registration software with optimal addressing the potentials of PC video card hardware greatly improves the speed of multi spectral image registration.

Fundus camera imaging is widely used to document disorders such as diabetic retinopathy and macular degeneration. Multiple retinal images can be combined together through a procedure known as mosaicing to form an image with a larger field of view. Mosaicing typically requires multiple pairwise registrations of partially overlapped images. We describe a new method for pairwise retinal image registration. The proposed method is unique in that the radial distortion due to image acquisition is corrected prior to the geometric transformation. Vessel lines are detected using the Hessian operator and are used as input features to the registration. Since the overlapping region is typically small in a retinal image pair, only a few correspondences are available, thus limiting the applicable model to an afine transform at best. To recover the distortion due to curved-surface of retina and lens optics, a combined approach of an afine model with a radial distortion correction is proposed. The parameters of the image acquisition and radial distortion models are estimated during an optimization step that uses Powell's method driven by the vessel line distance. Experimental results using 20 pairs of green channel images acquired from three subjects with a fundus camera confirmed that the afine model with distortion correction could register retinal image pairs to within 1.88±0.35 pixels accuracy (mean ± standard deviation) assessed by vessel line error, which is 17% better than the afine-only approach. Because the proposed method needs only two correspondences, it can be applied to obtain good registration accuracy even in the case of small overlap between retinal image pairs.

In biomedical research, there is an increased need for reconstruction of large soft tissue volumes (e.g. whole organs) at the microscopic scale from images obtained using laser scanning microscopy (LSM) with fluorescent dyes targeting selected cellular features. However, LSM allows reconstruction of volumes not exceeding a few hundred ìm in size and most LSM procedures require physical sectioning of soft tissue resulting in tissue deformation. Micro-CT (&mgr;CT) can provide deformation free tomographic image of the whole tissue volume before sectioning. Even though, the spatial resolution of &mgr;CT is around 5 &mgr;m and its contrast resolution is poor, it could provide information on external and internal interfaces of the investigated volume and therefore could be used as a template in the volume reconstruction from a very large number of LSM images. Here we present a method for accurate 3D reconstruction of the murine heart from large number of images obtained using confocal LSM. The volume is reconstructed in the following steps: (i) Montage synthesis of individual LSM images to form a set of aligned optical planes within given physical section; (ii) Image enhancement and segmentation to correct for non-uniform illumination and noise; (iii) Volume matching of a synthesized physical section to a corresponding sub-volume of &mgr;CT; (iv) Affine registration of the physical section to the selected &mgr;CT sub-volume. We observe correct gross alignment of the physical sections. However, many sections still exhibit local misalignment that could be only corrected via local nonrigid registration to &mgr;CT template and we plan to do it in the future.

Image registration is always a trade-off between accuracy and speed. Looking towards clinical scenarios the time for bringing two or more images into registration should be around a few seconds only. We present a new scheme for subsampling 3D-image data to allow for efficient computation of cost functions in intensity-based image registration. Starting from an arbitrary center point voxels are sampled along scan lines which do radially extend from the center point. We analyzed the characteristics of different cost functions computed on the sub-sampled data and compared them to known cost functions with respect to local optima. Results show the cost functions are smooth and give high peaks at the expected optima. Furthermore we investigated capture range of cost functions computed under the new subsampling scheme. Capture range was remarkably better for the new scheme compared to metrics using all voxels or different subsampling schemes and high registration accuracy was achieved as well. The most important result is the improvement in terms of speed making this scheme very interesting for clinical scenarios. We conclude using the new subsampling scheme intensity-based 3D image registration can be performed much faster than using other approaches while maintaining high accuracy. A variety of different extensions of the new approach is conceivable, e.g. non-regular distribution of the scan lines or not to let the scan lines start from a center point only, but from the surface of an organ model for example.

An established challenge in the field of image analysis has been the registration of images having a large initial misalignment. For example in chemo and Radiation Therapy Planning (RTP), there is often a need to register an image delineating a specific anatomy (usually in the surgery position) with that of a whole body image (obtained preoperatively). In such a scenario, there is room for a large misalignment between the two images that are required to be aligned. Large misalignments are traditionally handled in two ways: 1) Semi-automatically with a user initialization or 2) With the help of the origin fields in the image header. The first approach is user dependant and the second method can be used only if the two images are obtained from the same scanner with consistent origins. Our methodology extends a typical registration framework by selecting components that are capable of searching a large parameter space without settling on local optima. We have used an optimizer that is based on an Evolutionary Scheme along with an information theory based similarity metric that can address these needs. The attempt in this study is to convert a large misalignment problem to a small misalignment problem that can then be handled using application specific registration algorithms. Further improvements along local areas can be obtained by subjecting the image to a non-rigid transformation. We have successfully registered the following pairs of images without any user initialization: CTAC - simCT (neuro, lungs); MRPET/ CT (neuro, liver); T2-SPGR (neuro).

A number of intensity and feature based methods have been proposed for 3D to 2D registration. However, for multimodal 3D/2D registration of MR and X-ray images, only hybrid and reconstruction-based methods were shown to be feasible. In this paper we optimize the extraction of features in the form of bone edge gradients, which were proposed for 3D/2D registration of MR and X-ray images. The assumption behind such multimodal registration is that the extracted gradients in 2D X-ray images match well to the corresponding gradients extracted in 3D MR images. However, since MRI and X-rays are fundamentally different modalities, the corresponding bone edge gradients may not appear in the same position and the the above-mentioned assumption may thus not be valid. To test the validity of this assumption, we optimized the extraction of bone edges in 3D MR and also in CT images for the registration to 2D X-ray images. The extracted bone edges were systematically displaced in the direction of their gradients, i.e. in the direction of the normal to the bone surface, and corresponding effects on the accuracy and convergence of 3D/2D registration were evaluated. The evaluation was performed on two different sets of MR, CT and X-ray images of spine phantoms with known gold standard, first consisting of five and the other of eight vertebrae. The results showed that a better registration can be obtained if bone edges in MR images are optimized for each application-specific MR acquisition protocol.

We evaluate a novel method for fully automated rigid registration of 2D manifolds in 3D space based on distance maps, the Gibbs sampler and Iterated Conditional Modes (ICM). The method is tested against the ICP considered as the gold standard for automated rigid registration. Furthermore, the influence of different norms and sampling point densities is evaluated. The performance of the two methods has been evaluated on data consisting of 178 scanned ear impressions taken from the right ear. To quantify the difference of the two methods we calculate the registration cost and the mean point to point distance. T-test for common mean are used to determine the performance of the two methods (supported by a Wilcoxon signed rank test). The performance influence of sampling density, sampling quantity, and norms is analyzed using a similar method.

Symptoms of neurodegenerative pathologies such as Parkinson's disease can be relieved through Deep Brain Stimulation. This neurosurgical technique relies on high precision positioning of electrodes in specific areas of the basal ganglia and the thalamus. These subcortical anatomical targets must be located at pre-operative stage, from a set of MRI acquired under stereotactic conditions. In order to assist surgical planning, we designed a semi-automated image analysis process for extracting anatomical areas of interest. Complementary information, provided by both patient's data and expert knowledge, is represented as fuzzy membership maps, which are then fused by means of suitable possibilistic operators in order to achieve the segmentation of targets. More specifically, theoretical prior knowledge on brain anatomy is modelled within a 'virtual atlas' organised as a spatial graph: a list of vertices linked by edges, where each vertex represents an anatomical structure of interest and contains relevant information such as tissue composition, whereas each edge represents a spatial relationship between two structures, such as their relative directions. The model is built using heterogeneous sources of information such as qualitative descriptions from the expert, or quantitative information from prelabelled images. For each patient, tissue membership maps are extracted from MR data through a classification step. Prior model and patient's data are then matched by using a research algorithm (or 'strategy') which simultaneously computes an estimation of the location of every structures. The method was tested on 10 clinical images, with promising results. Location and segmentation results were statistically assessed, opening perspectives for enhancements.

During intensity-based 2D/3D registration of 3D CAD models of knee implant components to a calibrated single-plane fluoroscopy image, a similarity measure between the fluoroscopy and a rendering of the model onto the image plane is maximized w.r.t. the 3D pose parameters of the model. This work focuses on a robust strategy for finding this maximum by extending the standard Powell optimization algorithm with problem-specific knowledge. A combination of feature-based and intensity-based methods is proposed. At each iteration of the optimization process, feature information is used to compute an additional search direction along which the image-based similarity measure is maximized. Hence, the advantages of intensity-based registration (accuracy) and feature-based registration (robustness) are combined. The proposed method is compared with the standard Powell optimization strategy, using an image-based similarity measure only, on simulated fluoroscopy images. It is shown that the proposed method generally has higher accuracy, more robustness and a better convergence behavior. Although introduced for the registration of 3D CAD models of knee implant components to single-plane fluoroscopy images, the optimization strategy is easily extendible and applicable to other 2D/3D registration applications.

Non-rigid multi-modal registration of images/volumes is becoming increasingly necessary in many medical settings. While efficient registration algorithms have been published, the speed of the solutions is a problem in clinical applications. Harnessing the computational power of graphics processing unit (GPU) for general purpose computations has become increasingly popular in order to speed up algorithms further, but the algorithms have to be adapted to the data-parallel, streaming model of the GPU. This paper describes the implementation of a non-rigid, multi-modal registration using mutual information and the Kullback-Leibler divergence between observed and learned joint intensity distributions. The entire registration process is implemented on the GPU, including a GPU-friendly computation of two-dimensional histograms using vertex texture fetches as well as an implementation of recursive Gaussian filtering on the GPU. Since the computation is performed on the GPU, interactive visualization of the registration process can be done without bus transfer between main memory and video memory. This allows the user to observe the registration process and to evaluate the result more easily. Two hybrid approaches distributing the computation between the GPU and CPU are discussed. The first approach uses the CPU for lower resolutions and the GPU for higher resolutions, the second approach uses the GPU to compute a first approximation to the registration that is used as starting point for registration on the CPU using double-precision. The results of the CPU implementation are compared to the different approaches using the GPU regarding speed as well as image quality. The GPU performs up to 5 times faster per iteration than the CPU implementation.

Registration of medical images (intra- or multi-modality) is the first step before any analysis is performed. The analysis includes treatment monitoring, diagnosis, volumetric measurements or classification to mention a few. While pairwise registration, i.e., aligning a floating image to a fixed reference, is straightforward, it is not immediately clear what cost measures could be exploited for the groupwise alignment of several images (possibly multimodal) simultaneously. Recently however there has been increasing interest in this problem applied to atlas construction, statistical shape modeling, or simply joint alignment of images to get a consistent correspondence of voxels across all images based on a single cost measure. The aim of this paper is twofold, a) propose a cost function - alpha mutual information computed using entropic graphs that is a natural extension to Shannon mutual information for pairwise registration and b) compare its performance with the pairwise registration of the image set. We show that this measure can be reliably used to jointly align several images to a common reference. We also test its robustness by comparing registration errors for the registration process repeated at varying noise levels. In our experiments we used simulated data, applying different B-spline based geometric transformations to the same image and adding independent filtered Gaussian noise to each image. Non-rigid registration was employed with Thin Plate Splines(TPS) as the geometric interpolant.

We introduce two image alignment measures using Earth Mover's Distance (EMD) as a metric on the space of joint intensity distributions. Our first approach consists of computing EMD between a joint distribution and the product of its marginals. This yields a measure of statistical dependence comparable to Mutual Information, a criterion widely used for multimodal image registration. When a-priori knowledge is available, we also propose to compute EMD between the observed distribution and a joint distribution estimated from pairs of pre-aligned images. EMD is a cross-bin dissimilarity function and generally offers a generalization ability which is superior to previously proposed metrics, such as Kullback-Leibler divergence. Computing EMD amounts to solving an optimal mass transport problem whose solution can be very efficiently obtained using an algorithm recently proposed by Ling and Okada.¹⁰ We performed a preliminary experimental evaluation of this approach with real and simulated MR images. Our results show that EMD-based measures can be efficiently applied to rigid registration tasks.

One of the future-oriented areas of medical image processing is to develop fast and exact algorithms for image registration. By joining multi-modal images we are able to compensate the disadvantages of one imaging modality with the advantages of another modality. For instance, a Computed Tomography (CT) image containing the anatomy can be combined with metabolic information of a Positron Emission Tomography (PET) image. It is quite conceivable that a patient will not have the same position in both imaging systems. Furthermore some regions for instance in the abdomen can vary in shape and position due to different filling of the rectum. So a multi-modal image registration is needed to calculate a deformation field for one image in order to maximize the similarity between the two images, described by a so-called distance measure. In this work, we present a method to adapt a multi-modal distance measure, here mutual information (MI), with weighting masks. These masks are used to enhance relevant image structures and suppress image regions which otherwise would disturb the registration process. The performance of our method is tested on phantom data and real medical images.

Accurate registration of diagnosis and treatment images is a critical factor for the success of radiotherapy. This study presents a feature-based image registration algorithm that uses a branch- and-bound method to search the space of possible transformations, as well as a Hausdorff distance metric to evaluate their quality. This distance is computed in the space of responses to a circular Gabor filter, in which, for each point of interest in both reference and subject images, a vector of complex responses to different Gabor kernels is computed. Each kernel is generated using different frequencies and variances of the Gabor function, which determines correspondent regions in the images to be registered, by virtue of its rotation invariance characteristics. Responses to circular Gabor filters have also been reported in literature as a successful tool for image classification; and in this particular application we utilize them for patient positioning in cranial radiotherapy. For test purposes, we use 2D portal images acquired with an electronic portal imaging device (EPID). Our method presents EPID-EPID registrations errors under 0.2 mm for translations and 0.05 deg for rotations (subpixel accuracy). We are using fiducial marker registration as the ground truth for comparisons. Registration times average 2.70 seconds based on 1400 feature points using a 1.4 GHz processor.

Patient setup error is one of the major causes of tumor position uncertainty in radiotherapy for extracranial targets, which can result in a decreased radiation dose to the tumor and an increased dose to the normal tissues. Therefore, it is a common practice to verify the patient setup accuracy by comparing portal images with a digitally reconstructed radiograph (DRR) reference image. This paper proposes a practical method of portal image and DRR fusion for patient setup verification. As a result of the mean intensity difference between the inside and outside of the actual radiation region in the portal image, the image fusion in this work is fulfilled by applying an image registration process to the contents inside or outside of the actual radiation region in the portal image and the relevant contents that are extracted, accordingly, from the DRR image. The image fusion can also be fulfilled statistically by applying two separate image registration processes to the inside and outside of the actual radiation regions. To segment the image registration contents, automatic or semiautomatic region delineation schemes are employed that aim at minimizing users' operation burden, while at the same time maximizing the use of human intelligence. To achieve an accurate and fast delineation, this paper proposes using adaptive weight in the conventional level-set contour-finding algorithm for the automatic delineation scheme, as well as the use of adaptive banding in the conventional Intelligent Scissors algorithm for the semiautomatic delineation scheme.

Identifying the three-dimensional content of non-small cell lung cancer tumors is a vital step in the pursuit of understanding cancer growth, development and response to treatment. The majority of non-small cell lung cancer tumors are histologically heterogeneous, and consist of the malignant tumor cells, necrotic tumor cells, fibroblastic stromal tissue, and inflammation. Geometric and tissue density heterogeneity are utilized in computed tomography (CT) representations of lung tumors for distinguishing between malignant and benign nodules. However, the correlation between radiolographical heterogeneity and corresponding histological content has been limited. In this study, a multimodality dataset of human lung cancer is established, enabling the direct comparison between histologically identified tissue content and micro-CT representation. Registration of these two datasets is achieved through the incorporation of a large scale, serial microscopy dataset. This dataset serves as the basis for the rigid and non-rigid registrations required to align the radiological and histological data. The resulting comprehensive, three-dimensional dataset includes radio-density, color and cellular content of a given lung tumor. Using the registered datasets, neural network classification is applied to determine a statistical separation between cancerous and non-cancerous tumor regions in micro-CT.

Active Shape Models (ASM) have been applied to various segmentation tasks in medical imaging, most successfully in 2D segmentation of objects that have a fairly consistent shape. However, several difficulties arise when extending 2D ASM to 3D: (1) difficulty in 3D labeling, (2) the requirement of a large number of training samples, (3) the challenging problem of landmark correspondence in 3D, (4) inefficient initialization and optimization in 3D. This paper addresses the 3D segmentation problem by using a small number of effective 2D statistical models called oriented ASM (OASM). We demonstrate that a small number of 2D OASM models, which are derived from a chunk of a contiguous set of slices, are sufficient to capture the shape variation between slices and individual objects. Each model can be matched rapidly to a new slice by using the OASM algorithm¹. Our experiments in segmenting breast and bone of the foot in MR images indicate the following: (1) The accuracy of segmentation via our method is much better than that of 2DASM-based segmentation methods.² (2) Far fewer landmarks are required compared with thousands of landmarks needed in true 3D ASM. Therefore, far fewer training samples are required to capture details. (3) Our method is computationally slightly more expensive than the 2D method2 owing to its 2 level dynamic programming (2LDP) algorithm.

We developed an algorithm based on a rule-based threshold framework to segment the coronary arteries from angiographic computed tomography (CTA) data. Computerized segmentation of the coronary arteries is a challenging procedure due to the presence of diverse anatomical structures surrounding the heart on cardiac CTA data. The proposed algorithm incorporates various levels of image processing and organ information including region, connectivity and morphology operations. It consists of three successive stages. The first stage involves the extraction of the three-dimensional scaffold of the heart envelope. This stage is semiautomatic requiring a reader to review the CTA scans and manually select points along the heart envelope in slices. These points are further processed using a surface spline-fitting technique to automatically generate the heart envelope. The second stage consists of segmenting the left heart chambers and coronary arteries using grayscale threshold, size and connectivity criteria. This is followed by applying morphology operations to further detach the left and right coronary arteries from the aorta. In the final stage, the 3D vessel tree is reconstructed and labeled using an Isolated Connected Threshold technique. The algorithm was developed and tested on a patient coronary artery CTA that was graciously shared by the Department of Radiology of the Massachusetts General Hospital. The test showed that our method constantly segmented the vessels above 79% of the maximum gray-level and automatically extracted 55 of the 58 coronary segments that can be seen on the CTA scan by a reader. These results are an encouraging step toward our objective of generating high resolution models of the male and female heart that will be subsequently used as phantoms for medical imaging system optimization studies.

The segmentation of the subcortical structures of the brain is required for many forms of quantitative neuroanatomic analysis. The volumetric and shape parameters of structures such as lateral ventricles, putamen, caudate, hippocampus, pallidus and amygdala are employed to characterize a disease or its evolution. This paper presents a fully automatic segmentation of these structures via a non-rigid registration of a probabilistic atlas prior and alongside a comprehensive validation. Our approach is based on an unbiased diffeomorphic atlas with probabilistic spatial priors built from a training set of MR images with corresponding manual segmentations. The atlas building computes an average image along with transformation fields mapping each training case to the average image. These transformation fields are applied to the manually segmented structures of each case in order to obtain a probabilistic map on the atlas. When applying the atlas for automatic structural segmentation, an MR image is first intensity inhomogeneity corrected, skull stripped and intensity calibrated to the atlas. Then the atlas image is registered to the image using an affine followed by a deformable registration matching the gray level intensity. Finally, the registration transformation is applied to the probabilistic maps of each structures, which are then thresholded at 0.5 probability. Using manual segmentations for comparison, measures of volumetric differences show high correlation with our results. Furthermore, the dice coefficient, which quantifies the volumetric overlap, is higher than 62% for all structures and is close to 80% for basal ganglia. The intraclass correlation coefficient computed on these same datasets shows a good inter-method correlation of the volumetric measurements. Using a dataset of a single patient scanned 10 times on 5 different scanners, reliability is shown with a coefficient of variance of less than 2 percents over the whole dataset. Overall, these validation and reliability studies show that our method accurately and reliably segments almost all structures. Only the hippocampus and amygdala segmentations exhibit relative low correlation with the manual segmentation in at least one of the validation studies, whereas they still show appropriate dice overlap coefficients.

A multi-resolution, parallel approach to retinal blood vessel detection has been introduced that can also be used as a discriminant for fovea detection. Localized adaptive thresholding and a multi-resolution, multi-window Radon transform (RT) are utilized to detect the retinal vascular system. Multi-window parameter transforms are intrinsically parallel and offer increased performance over conventional transforms. Large vessels are extracted in low-resolution mode, whereas minor vessels are extracted in high-resolution mode further increasing computational efficiency. The image is adaptively thresholded and then the multi-window RT is applied at the different resolution levels. Results from each level are combined and morphologically processed to improve final performance. A systematic approach has been implemented to perform fovea detection. The algorithm relies on a probabilistic method to perform initial segmentation. The intensity image is re-mapped into probability space to detect areas with low-probability of occurrence. Intensity and probability information are coupled to produce a binary image that contains potential fovea candidates. The candidates are discriminated based upon their location within the blood vessel network.

Many neuroimaging studies are applied to primates as pathologies and environmental exposures can be studied in well-controlled settings and environment. In this work, we present a framework for both the semi-automatic creation of a rhesus monkey atlas and a fully automatic segmentation of brain tissue and lobar parcellation. We determine the atlas from training images by iterative, joint deformable registration into an unbiased average image. On this atlas, probabilistic tissue maps and a lobar parcellation. The atlas is then applied via affine, followed by deformable registration. The affinely transformed atlas is employed for a joint T1/T2 based tissue classification. The deformed atlas parcellation masks the tissue segmentations to define the parcellation. Other regional definitions on the atlas can also straightforwardly be used as segmentation. We successfully built average atlas images for the T1 and T2 datasets using a developmental training datasets of 18 cases aged 16-34 months. The atlas clearly exhibits an enhanced signal-to-noise ratio compared to the original images. The results further show that the cortical folding variability in our data is highly limited. Our segmentation and parcellation procedure was successfully re-applied to all training images, as well as applied to over 100 additional images. The deformable registration was able to identify corresponding cortical sulcal borders accurately. Even though the individual methods used in this segmentation framework have been applied before on human data, their combination is novel, as is their adaptation and application to rhesus monkey MRI data. The reduced variability present in the primate data results in a segmentation pipeline that exhibits high stability and anatomical accuracy.

This work proposes an algorithm for segmentation and tracking of human sperm. The algorithm analyzes video sequences containing multiple moving sperms and produces video segmentation maps and moving objects trajectories. Sperm trajectories analysis is widely used in computer-aided sperm analysis (CASA) systems. Several researches show that CASA systems face a problem when dealing with the "actual" or "perceived" collisions of sperms. The proposed algorithm reduces the probability of wrong trajectory construction related to collisions. We represent the video data using a 4-dimensional model containing spatial and temporal coordinates and the direction of optical flow vectors. The video sequence is divided into a succession of overlapping subsequences. The video data of each subsequence is grouped in the feature domain using the mean shift procedure. We identify clusters corresponding to moving objects in each subsequence. The complete trajectories are reconstructed by matching clusters that are most likely to represent that same object in adjacent subsequences. The clusters are matched using heuristics which are based on cluster overlaps, and by solving a specially formulated linear assignment problem. Tracking results are evaluated for different video sequences containing different types of motions and collisions.

The full-width at half-max (FWHM) criterion is often used for both manual and automatic quantification of the vessel diameter in medical images. The FWHMcriterion is easy to understand and it can be implemented with low computational cost. However, it is well known that the FWHM criterion can give an over- and underestimation of the vessel diameter. In this paper, we propose a simple and original method to create an unbiased estimation of the vessel diameter based on the FWHM criterion and we compared the robustness to noise of several edge detectors. The quantitative results of our experiments show that the proposed method is more accurate and precise than other (more complex) edge detectors, even for small vessels.

Bronchopulmonary segments are valuable as they give more accurate localization than lung lobes. Traditionally, determining the segments requires segmentation and identification of segmental bronchi, which, in turn, require volumetric imaging data. In this paper, we present a method for approximating the bronchopulmonary segments for sparse data by effectively using an anatomical atlas. The atlas is constructed from a volumetric data and contains accurate information about bronchopulmonary segments. A new ray-tracing based image registration is used for transferring the information from the atlas to a query image. Results show that the method is able to approximate the segments on sparse HRCT data with slice gap up to 25 millimeters.

The study deals with the challenging task of automatic segmentation of MR brain images with multiple sclerosis lesions (MSL). Multi-Channel data is used, including "fast fluid attenuated inversion recovery" (fast FLAIR or FF), and statistical modeling tools are developed, in order to improve cerebrospinal fluid (CSF) classification and to detect MSL. Two new concepts are proposed for use within an EM framework. The first concept is the integration of prior knowledge as it relates to tissue behavior in different MRI modalities, with special attention given to the FF modality. The second concept deals with running the algorithm on a subset of the input that is most likely to be noise- and artifact-free data. This enables a more reliable learning of the Gaussian mixture model (GMM) parameters for brain tissue statistics. The proposed method focuses on the problematic CSF intensity distribution, which is a key to improved overall segmentation and lesion detection. A level-set based active contour stage is performed for lesion delineation, using gradient and shape properties combined with previously learned region intensity statistics. In the proposed scheme there is no need for preregistration of an atlas, a common characteristic in brain segmentation schemes. Experimental results on real data are presented.

Patients suffering from a heart valve deficiency are often treated by replacing the valve with an artificial or biological implant. In case of biological implants, the use of porcine heart valves is common. Quality assessment and inspection methods are mandatory to supply the patients (and also medical research) with only the best such xenograft implants thus reducing the number of follow-up surgeries to replace worn-up valves. We describe an approach for automatic in-vitro evaluation of prosthetic heart valves in an artificial circulation system. We show how to measure the orifice area during a heart cycle to obtain an orifice curve. Different quality parameters are then estimated on such curves.

Skull is the anatomic landmark for patient set up of head radiation therapy. Skull is generally segmented from CT images because CT provides better definition of skull than MR imaging. In the mean time, radiation therapy is planned on MR images for soft tissue information. This study utilized a knowledge-guided active model (KAM) method to segmented skull on MR images in order to enable radiation therapy planning with MR images as the primary planning dataset. KAM utilized age-specific skull mesh models that segmented from CT images using a conditional region growing algorithm. Skull models were transformed to given MR images using an affine registration algorithm based on normalized mutual information. The transformed mesh models actively located skull boundaries by minimizing their total energy. The preliminary validation was performed on MR and CT images from five patients. The KAM segmented skulls were compared with those segmented from CT images. The average image similarity (kappa index) was 0.57. The initial validation showed that it was promising to segment skulls directly on MR images using KAM.

Neurodegenerative and neurodevelopmental diseases demonstrate problems associated with brain maturation and aging. Automated methods to delineate brain structures of interest are required to analyze large amounts of imaging data like that being collected in several on going multi-center studies. We have previously reported on using artificial neural networks (ANN) to define subcortical brain structures including the thalamus (0.88), caudate (0.85) and the putamen (0.81). In this work, apriori probability information was generated using Thirion's demons registration algorithm. The input vector consisted of apriori probability, spherical coordinates, and an iris of surrounding signal intensity values. We have applied the support vector machine (SVM) machine learning algorithm to automatically segment subcortical and cerebellar regions using the same input vector information. SVM architecture was derived from the ANN framework. Training was completed using a radial-basis function kernel with gamma equal to 5.5. Training was performed using 15,000 vectors collected from 15 training images in approximately 10 minutes. The resulting support vectors were applied to delineate 10 images not part of the training set. Relative overlap calculated for the subcortical structures was 0.87 for the thalamus, 0.84 for the caudate, 0.84 for the putamen, and 0.72 for the hippocampus. Relative overlap for the cerebellar lobes ranged from 0.76 to 0.86. The reliability of the SVM based algorithm was similar to the inter-rater reliability between manual raters and can be achieved without rater intervention.

One aim of this work is to investigate the feasibility of using a hierarchy of models to describe diffusion tensor MRI data. Parsimonious model selection criteria are used to choose among different models of diffusion within tissue. Second, based on this information, we assess whether we can perform simultaneous tissue segmentation and classification. The proposed hierarchical framework used for parsimonious model selection is based on the F-test, adapted from Snedecor. Diffusion Magnetic Resonance Microscopy (MRM) provides near-microscopic resolution without relying on a sample's optical transparency for image formation. Diffusion MRM is a noninvasive imaging technique for quantitative analysis of intrinsic features of tissues. Thus, we propose using Diffusion MRM to characterize normal tissue structure in adult zebrafish, and possibly subtle anatomical or structural differences between normals and knockouts. Both numerical phantoms and diffusion weighted image (DWI) data obtained from adult zebrafish are used to test this model selection framework.

We have previously developed a breast boundary detection method by using a gradient-based method to search for the breast boundary (GBB). In this study, we developed a new dynamic multiple thresholding based breast boundary detection system (MTBB). The initial breast boundary (MTBB-Initial) is obtained based on the analysis of multiple thresholds on the image. The final breast boundary (MTBB-Final) is obtained based on the initial breast boundary and the gradient information from horizontal and the vertical Sobel filtering. In this way, it is possible to accurately segment the breast area from the background region. The accuracy of the breast boundary detection algorithm was evaluated by comparison with an experienced radiologist's manual segmentation using three performance metrics: the Hausdorff distance (HDist), the average minimum Euclidean distance (AMinDist), and the area overlap (AOM). It was found that 68%, 85%, and 90% of images have HDist errors less than 6 mm for GBB, MTBB-Initial, and MTBB-Final, respectively. Ninety-five percent, 96%, and 97% of the images have AMinDist errors less than 1.5 mm for GBB, MTBB-Initial, and MTBB-Final, respectively. Ninety-six percent, 97%, and 99% of the images have AOM values larger than 0.9 for GBB, MTBB-Initial, and MTBB-Final, respectively. It was found that the performance of the proposed method was improved in comparison to our previous method.

Automatic liver segmentation is still a challenging task due to the ambiguity of liver boundary and the complex context of nearby organs. In this paper, we propose a faster and more accurate way of liver segmentation in CT images with an enhanced level set method. The speed image for level-set propagation is smoothly generated by increasing number of iterations in anisotropic diffusion filtering. This prevents the level-set propagation from stopping in front of local minima, which prevails in liver CT images due to irregular intensity distributions of the interior liver region. The curvature term of shape modeling level-set method captures well the shape variations of the liver along the slice. Finally, rolling ball algorithm is applied for including enhanced vessels near the liver boundary. Our approach are tested and compared to manual segmentation results of eight CT scans with 5mm slice distance using the average distance and volume error. The average distance error between corresponding liver boundaries is 1.58 mm and the average volume error is 2.2%. The average processing time for the segmentation of each slice is 5.2 seconds, which is much faster than the conventional ones. Accurate and fast result of our method will expedite the next stage of liver volume quantification for liver transplantations.

We are investigating in vivo small animal imaging and analysis methods for the assessment of photodynamic therapy (PDT), an emerging therapeutic modality for cancer treatment. Multiple weighted MR images were acquired from tumor-bearing mice pre- and post-PDT and 24-hour after PDT. We developed an automatic image classification method to differentiate live, necrotic and intermediate tissues within the treated tumor on the MR images. We used a multiscale diffusion filter to process the MR images before classification. A multiscale fuzzy C-means (FCM) classification method was applied along the scales. The object function of the standard FCM was modified to allow multiscale classification processing where the result from a coarse scale is used to supervise the classification in the next scale. The multiscale fuzzy C-means (MFCM) method takes noise levels and partial volume effects into the classification processing. The method was validated by simulated MR images with various noise levels. For simulated data, the classification method achieved 96.0 ± 1.1% overlap ratio. For real mouse MR images, the classification results of the treated tumors were validated by histologic images. The overlap ratios were 85.6 ± 5.1%, 82.4 ± 7.8% and 80.5 ± 10.2% for the live, necrotic, and intermediate tissues, respectively. The MR imaging and the MFCM classification methods may provide a useful tool for the assessment of the tumor response to photodynamic therapy in vivo.

Multidetector row CT, multiphase CT in particular, has been widely accepted as a sensitive imaging modality in the detection of liver cancer. Segmentation of liver from CT images is of great importance in terms of accurate detection of tumours, volume measurement, pre-surgical planning. The segmentation of liver, however, remains to be an unsolved problem due to the complicated nature of liver CT such as imaging noise, similar intensity to its adjacent structures and large variations of contrast kinetics and localised geometric features. The purpose of this paper is to present our newly developed algorithm aiming to tackle this problem. In our method, a CT image was first smoothed by geometric diffusion method; the smoothed image was segmented by thresholding operators. In order to gain optimal segmentation, a novel method was developed to choose threshold values based on both the anatomical knowledge and features of liver CT. Then morphological operators were applied to fill the holes in the generated binary image and to disconnect the liver from other unwanted adjoining structures. After this process, a so-called "2.5D region overlapping" filter was introduced to further remove unwanted regions. The resulting 3D region was regarded as the final segmentation of the liver region. This method was applied to venous phase CT data of 45 subjects (30 patient and 15 asymptomatic subjects). Our results show good agreement with the annotations delineated manually by radiologists and the overlapping ratio of volume is 87.7% on average and the correlation coefficient between them is 98.1%.

Prostate cancer is the most common cancer in men. Tissue extraction at different locations (biopsy) is the gold-standard for diagnosis of prostate cancer. These biopsies are commonly guided by transrectal ultrasound imaging (TRUS). Exact location of the extracted tissue within the gland is desired for more specific diagnosis and provides better therapy planning. While the orientation and the position of the needle within clinical TRUS image are limited, the appearing length and visibility of the needle varies strongly. Marker lines are present and tissue inhomogeneities and deflection artefacts may appear. Simple intensity, gradient oder edge-detecting based segmentation methods fail. Therefore a multivariate statistical classificator is implemented. The independent feature model is built by supervised learning using a set of manually segmented needles. The feature space is spanned by common binary object features as size and eccentricity as well as imaging-system dependent features like distance and orientation relative to the marker line. The object extraction is done by multi-step binarization of the region of interest. The ROI is automatically determined at the beginning of the segmentation and marker lines are removed from the images. The segmentation itself is realized by scale-invariant classification using maximum likelihood estimation and Mahalanobis distance as discriminator. The technique presented here could be successfully applied in 94% of 1835 TRUS images from 30 tissue extractions. It provides a robust method for biopsy needle localization in clinical prostate biopsy TRUS images.

Fully automatic segmentation on medical images often generates unreliable results so we must rely on semi-automatic methods that use both user input and boundary refinement to produce a more accurate result. In this paper, we present an improved livewire method for noisy regions of interest with low contrast boundaries. The first improvement is the adaptive search space, which minimizes the required search area for graph generation, and a directional graph searching which also speeds up the shortest path finding. The second improvement is an enhanced cost function to consider only the local maximum gradient within our search area, which prevents interference from objects we are not interested in. The third improvement is the on-the-fly training based on gradient histogram to prevent attraction of the contour to strong edges that are not part of the actual contour. We carried out tests between the original and our improved version of livewire. The segmentation was validated on phantom images and also against manual segmentation defined by experts on uterine leiomyomas MRI. Our results show that, on average, our method reduces the time to completion by 96% with improved accuracy up to 63%.

We describe an algorithm for evolving a level set on a non-planar manifold like the isosurface of a 3D object. The surface is represented by a triangular mesh and the feature that guides our level set via a speed function is its curvature. We overcome the difficulty of computing the gradient and curvature of the level set distance function on a non-planar, non-Cartesian mesh by performing the calculations locally in neighborhoods small enough to be considered planar. Moreover we use a least squares estimation of derivatives to replace finite differences and achieve better accuracy. The algorithm was motivated by our need to detect the ridge lines of vertebral bodies. The advantage of using level sets is that they are capable of producing a continuous ridge line despite noise and gaps in the ridge. We tested our algorithm on 40 vertebral bodies (80 ridge lines). 76 ridge lines showed no noticeable mistakes. The same set of parameters was used. For the remaining 4, we had to change the parameters of the speed function sigmoid to correct small under- or over-segmenting. To further test our algorithm we designed a synthetic surface with large curvature and to which we added noise and a ridge. The level set was able to evolve on the surface and stop at the ridge. Tests on synthetic cylinders with a ground truth ridge and to which we added noise indicate that the level set has good accuracy.

Segmentation of 3D data is one of the most challenging tasks in medical image analysis. While reliable automatic methods are typically preferred, their success is often hindered by poor image quality and significant variations in anatomy. Recent years have thus seen an increasing interest in the development of semi-automated segmentation methods that combine computational tools with intuitive, minimal user interaction. In an earlier work, we introduced a highly-automated technique for medical image segmentation, where a 3D extension of the traditional 2D Livewire was proposed. In this paper, we present an enhanced and more powerful 3D Livewire-based segmentation approach with new features designed to primarily enable the handling of complex object topologies that are common in biological structures. The point ordering algorithm we proposed earlier, which automatically pairs up seedpoints in 3D, is improved in this work such that multiple sets of points are allowed to simultaneously exist. Point sets can now be automatically merged and split to accommodate for the presence of concavities, protrusions, and non-spherical topologies. The robustness of the method is further improved by extending the 'turtle algorithm', presented earlier, by using a turtle-path pruning step. Tests on both synthetic and real medical images demonstrate the efficiency, reproducibility, accuracy, and robustness of the proposed approach. Among the examples illustrated is the segmentation of the left and right ventricles from a T1-weighted MRI scan, where an average task time reduction of 84.7% was achieved when compared to a user performing 2D Livewire segmentation on every slice.

It is still unclear whether periventricular and subcortical white matter lesions (WMLs) differ in etiology or clinical consequences. Studies addressing this issue would benefit from automated segmentation and localization of WMLs. Several papers have been published on WML segmentation in MR images. Automated localization however, has not been investigated as much. This work presents and evaluates a novel method to label segmented WMLs as periventricular and subcortical. The proposed technique combines tissue classification and registration-based segmentation to outline the ventricles in MRI brain data. The segmented lesions can then be labeled into periventricular WMLs and subcortical WMLs by applying region growing and morphological operations. The technique was tested on scans of 20 elderly subjects in which neuro-anatomy experts manually segmented WMLs. Localization accuracy was evaluated by comparing the results of the automated method with a manual localization. Similarity indices and volumetric intraclass correlations between the automated and the manual localization were 0.89 and 0.95 for periventricular WMLs and 0.64 and 0.89 for subcortical WMLs, respectively. We conclude that this automated method for WML localization performs well to excellent in comparison to the gold standard.

In this paper, we investigate several improvements to region-based level set algorithms in the context of segmenting x-ray CT data from pre-clinical imaging of small animal models. We incorporate a recently introduced signed distance preserving term into a region-based level set model and provide formulas for a semi-implicit finite difference implementation. We illustrate some pitfalls of topology preserving level sets and introduce the concept of connectivity preservation as a potential alternative. We illustrate the benefits of these improvements on phantom and real data.

We introduce a new model-based approach for segmenting and quantifying fluorescent bacteria in 3D microscopy live cell images. The approach is based on a new 3D superellipsoidal parametric intensity model, which is directly fitted to the image intensities within 3D regions-of-interest. Based on the fitting results, we can directly compute the total amount of intensity (fluorescence) of each cell. In addition, we introduce a method for automatic initialization of the model parameters, and we propose a method for simultaneously fitting clustered cells by using a superposition of 3D superellipsoids for model fitting. We demonstrate the applicability of our approach based on 3D synthetic and real 3D microscopy images.

Segmentation of an abnormal liver region based on CT or MR images is a crucial step in surgical planning. However, precisely carrying out this step remains a challenge due to either connectivities of the liver to other organs or the shape, internal texture, and homogeneity of liver that maybe extensively affected in case of liver diseases. Here, we propose a non-density based method for extracting the liver region containing tumor tissues by edge detection processing. False extracted regions are eliminated by a shape analysis method and thresholding processing. If the multi-phased images are available then the overall outcome of segmentation can be improved by subtracting two phase images, and the connectivities can be further eliminated by referring to the intensity on another phase image. Within an edge liver map, tumor candidates are identified by their different gray values relative to the liver. After elimination of the small and nonspherical over-extracted regions, the final liver region integrates the tumor region with the liver tissue. In our experiment, 40 cases of MDCT images were used and the result showed that our fully automatic method for the segmentation of liver region is effective and robust despite the presence of hepatic tumors within the liver.

The striatum is the input component of the basal ganglia from the cerebral cortex. It includes the caudate, putamen, and nucleus accumbens. Thus, the striatum is an important component in limbic frontal-subcortical circuitry and is believed to be relevant both for reward-guided behaviors and for the expression of psychosis. The dorsal striatum is composed of the caudate and putamen, both of which are further subdivided into pre- and post-commissural components. The ventral striatum (VS) is primarily composed of the nucleus accumbens. The striatum can be functionally divided into three broad regions: 1) a limbic; 2) a cognitive and 3) a sensor-motor region. The approximate corresponding anatomic subregions for these 3 functional regions are: 1) the VS; 2) the pre/post-commissural caudate and the pre-commissural putamen and 3) the post-commissural putamen. We believe assessing these subregions, separately, in disorders with limbic and cognitive impairment such as schizophrenia may yield more informative group differences in comparison with normal controls than prior parcellation strategies of the striatum such as assessing the caudate and putamen. The manual parcellation of the striatum into these subregions is currently defined using certain landmark points and geometric rules. Since identification of these areas is important to clinical research, a reliable and fast parcellation technique is required. Currently, only full manual parcellation using editing software is available; however, this technique is extremely time intensive. Previous work has shown successful application of heuristic rules into a semi-automatic platform1. We present here a semi-automatic algorithm which implements the rules currently used for manual parcellation of the striatum, but requires minimal user input and significantly reduces the time required for parcellation.

Microaneurysms (MAs) detection is a critical step in diabetic retinopathy screening, since MAs are the earliest visible warning of potential future problems. A variety of algorithms have been proposed for MAs detection in mass screening. Different methods have been proposed for MAs detection. The core technology for most of existing methods is based on a directional mathematical morphological operation called "Top-Hat" filter that requires multiple filtering operations at each pixel. Background structure, uneven illumination and noise often cause confusion between MAs and some non-MA structures and limits the applicability of the filter. In this paper, a novel detection framework based on edge directed inference is proposed for MAs detection. The candidate MA regions are first delineated from the edge map of a fundus image. Features measuring shape, brightness and contrast are extracted for each candidate MA region to better exclude false detection from true MAs. Algorithmic analysis and empirical evaluation reveal that the proposed edge directed inference outperforms the "Top-Hat" based algorithm in both detection accuracy and computational speed.

Image segmentation and delineation is at the heart of modern radiotherapy, where the aim is to deliver as high a radiation dose as possible to a cancerous target whilst sparing the surrounding healthy tissues. This, of course, requires that a radiation oncologist dictates both where the tumour and any nearby critical organs are located. As well as in treatment planning, delineation is of vital importance in image guided radiotherapy (IGRT): organ motion studies demand that features across image databases are accurately segmented, whilst if on-line adaptive IGRT is to become a reality, speedy and correct target identification is a necessity. Recently, much work has been put into the development of automatic and semi-automatic segmentation tools, often using prior knowledge to constrain some grey level, or derivative thereof, interrogation algorithm. It is hoped that such techniques can be applied to organ at risk and tumour segmentation in radiotherapy. In this work, however, we make the assumption that grey levels do not necessarily determine a tumour's extent, especially in CT where the attenuation coefficient can often vary little between cancerous and normal tissue. In this context we present an algorithm that generates a discontinuity free delineation surface driven by user placed, evidence based support points. In regions of sparse user supplied information, prior knowledge, in the form of a statistical shape model, provides guidance. A small case study is used to illustrate the method. Multiple observers (between 3 and 7) used both the presented tool and a commercial manual contouring package to delineate the bladder on a serially imaged (10 cone beam CT volumes ) prostate patient. A previously presented shape analysis technique is used to quantitatively compare the observer variability.

This paper aims at automatically measuring the extent of calcified plaques in the lumbar aorta from standard radiographs. Calcifications in the abdominal aorta are an important predictor for future cardiovascular morbidity and mortality. Accurate and reproducible measurement of the amount of calcified deposit in the aorta is therefore of great value in disease diagnosis and prognosis, treatment planning, and the study of drug effects. We propose a two-step approach in which first the calcifications are detected by an iterative statistical pixel classification scheme combined with aorta shape model optimization. Subsequently, the detected calcified pixels are used as the initialization for an inpainting based segmentation. We present results on synthetic images from the inpainting based segmentation as well as results on several X-ray images based on the two-steps approach.

Hepatic vessel trees are the key structures in the liver. Knowledge of the hepatic vessel trees is important for liver surgery planning and hepatic disease diagnosis such as portal hypertension. However, hepatic vessels cannot be easily distinguished from other liver tissues in non-contrast CT images. Automated segmentation of hepatic vessels in non-contrast CT images is a challenging issue. In this paper, an approach for automated segmentation of hepatic vessels trees in non-contrast X-ray CT images is proposed. Enhancement of hepatic vessels is performed using two techniques: (1) histogram transformation based on a Gaussian window function; (2) multi-scale line filtering based on eigenvalues of Hessian matrix. After the enhancement of hepatic vessels, candidate of hepatic vessels are extracted by thresholding. Small connected regions of size less than 100 voxels are considered as false-positives and are removed from the process. This approach is applied to 20 cases of non-contrast CT images. Hepatic vessel trees segmented from the contrast-enhanced CT images of the same patient are used as the ground truth in evaluating the performance of the proposed segmentation method. Results show that the proposed method can enhance and segment the hepatic vessel regions in non-contrast CT images correctly.

While there are many publicly available software packages for medical image processing, making them available to end users in clinical and research labs remains non-trivial. An even more challenging task is to mix these packages to form pipelines that meet specific needs seamlessly, because each piece of software usually has its own input/output formats, parameter sets, and so on. To address these issues, we are building WHIPPET (Washington Heterogeneous Image Processing Pipeline EnvironmenT), a collaborative platform for integrating image analysis tools from different sources. The central idea is to develop a set of Python scripts which glue the different packages together and make it possible to connect them in processing pipelines. To achieve this, an analysis is carried out for each candidate package for WHIPPET, describing input/output formats, parameters, ROI description methods, scripting and extensibility and classifying its compatibility with other WHIPPET components as image file level, scripting level, function extension level, or source code level. We then identify components that can be connected in a pipeline directly via image format conversion. We set up a TWiki server for web-based collaboration so that component analysis and task request can be performed online, as well as project tracking, knowledge base management, and technical support. Currently WHIPPET includes the FSL, MIPAV, FreeSurfer, BrainSuite, Measure, DTIQuery, and 3D Slicer software packages, and is expanding. Users have identified several needed task modules and we report on their implementation.

This study is conducted to assess the influence of various CT imaging parameters and airway obliquity, such as reconstruction kernel, field of view, slice thickness, and obliquity of airway on automatic measurement of airway wall thickness with FWHM method and physical phantom. The phantom, consists of 11 poly-acryl tubes with various inner lumen diameters and thickness, was used in this study. The measured density of the wall was 150HU. The airspace outside of tube was filled with poly-urethane foam, whose density was -900HU, which is similar density of emphysema region. CT images, obtained with MDCT (Sensation 16, Siemens), was reconstructed with various reconstruction kernel (B10f, B30f, B50f, B70f and B80f), different field of views (180mm, 270mm, 360mm), and different thicknesses (0.75, 1, and 2 mm). The phantom was scanned at various oblique angles (0, 30, 45, 60 degree). Using in-house airway measurement software, central axis of oblique airway was determined by 3D thinning algorithm and CT image perpendicular to the axis was reconstructed. The luminal area, outer boundary, and wall thickness was measured by FWHM method at each image. Actual dimension of each tube and measured CT values on each CT data set was compared. Sharper reconstruction kernel, thicker image thickness, and larger oblique angle of airway axis results in decrease of measured wall thickness. There was internal interaction between imaging parameters and obliquity of airway on the accuracy of measurement. There was a threshold point of 1-mm wall thickness, below which the measurement failed to represent the change of real thickness. Even using the smaller FOV, the accuracy was not improved. Usage of standard kernel (B50f) and 0.75mm thickness results in the most accurate measurement results, which is independent of obliquity of airway. (Mean error: 0 Degree 0.067±0.05mm, 30 Degree 0.076±0.09, 45 Degree 0.074±0.09, 60 Degree 0.091±0.09). In this imaging parameters, there was no significant difference (paired t-test : p > 0.05) between actual measurement and each oblique angle measurement. The accuracy of airway wall measurement was strongly influenced by imaging parameters and obliquity of airway. For the accurate measurement, independent of obliquity, we recommend the CT images reconstructed with 0.75mm slice thickness and B50f or B30f with sharpening filter.

Tomographic reconstructions, which are generally gray-scale images, are often segmented as to extract quantitative information, such as the shape or volume of image objects. At present, segmentation is usually performed by thresholding. However, the process of threshold selection is somewhat arbitrary and requires human interaction. In this paper, we present an algorithmic approach for automatically selecting the segmentation thresholds by using the available tomographic projection data. Assuming that each material (i.e., tissue type) in the sample has a characteristic, approximately constant gray value, thresholds are computed for which the segmented image corresponds optimally with the projection data.

In this paper, we present a method that detects lesions in two-dimensional (2D) cross-sectional brain images. By calculating the major and minor axes of the brain, we calculate an estimate of the background, without any a priori information, to use in inverse filtering. Shape saliency computed by a Gabor filter bank is used to further refine the results of the inverse filtering. The proposed algorithm was tested on different images of "The Whole Brain Atlas" database. The experimental results have produced 93% classification accuracy in processing 100 arbitrary images, representing different kinds of brain lesion.

Quantification of cervical carotid geometry may facilitate improved clinical decision making and scientific discovery. We set out to evaluate the ability of InsightSNAP (ITK-SNAP), an open-source segmentation program for 3D medical images (http://www.itksnap.org, version 1.4), to semi-automatically segment internal carotid arteries. A sample of five individuals (three normal volunteers, and two diseased patients) were imaged with an MR exam consisting of a MOTSA TOF MRA image volume and multiple black blood images acquired with different contrast weightings. Comparisons were made to a manual segmentation created during simultaneous evaluation of the MOTSA image and the various black blood images (typically PD-weighted, T1-weighted, and T2-weighted). These individuals were selected as a training set to determine acceptable parameters for ITK-SNAP's semi-automatic level sets segmentation method. The conclusion from this training set was that the initial thresholding (assigning probabilities to the intensities of image pixels) in the image pre-processing step was most important to obtaining an acceptable segmentation. Unfortunately no consistent trends emerged in how this threshold should be chosen. Figures of percent over- and under-segmentation were computed as a means of comparing the hand segmented and semi-automatically segmented internal carotids. Overall the under-segmentation by ITK-SNAP (voxels included in the manual segmentation but not in the semiautomated segmentation) was 10.94% ± 6.35% while the over-segmentation (voxels excluded in the manual segmentation but included in the semi-automated segmentation) was 8.16% ± 4.40% defined by reference to the total number of voxels included in the manual segmentation.

To segment low density lung regions in chest CT images, most of methods use the difference in gray-level value of pixels. However, radiodense pulmonary vessels and pleural nodules that contact with the surrounding anatomy are often excluded from the segmentation result. To smooth lung boundary segmented by gray-level processing in chest CT images, we propose a new method using scan line search. Our method consists of three main steps. First, lung boundary is extracted by our automatic segmentation method. Second, segmented lung contour is smoothed in each axial CT slice. We propose a scan line search to track the points on lung contour and find rapidly changing curvature efficiently. Finally, to provide consistent appearance between lung contours in adjacent axial slices, 2D closing in coronal plane is applied within pre-defined subvolume. Our method has been applied for performance evaluation with the aspects of visual inspection, accuracy and processing time. The results of our method show that the smoothness of lung contour was considerably increased by compensating for pulmonary vessels and pleural nodules.

Delineation of tumor and organs at risk on each phase of 4D CT images is an essential step in adaptive radiotherapy planning. Manual contouring of the large amount of data is time-consuming and impractical. (Semi-) automated methods typically rely on deformable image registration techniques to automatically map the manual contours drawn in one image to all the other phases in order to get complete 4D contouring, a procedure known as automatic re-contouring. Disadvantages of such approaches are that the manual contouring information is not used in the registration process and the whole volume registration is highly inefficient. In this work, we formulate the automatic re-contouring in a deformable surface model framework, which effectively restricts the computation to a lower dimensional space. The proposed framework was inspired by the morphing active contour model proposed by Bertalmio et al. [1], but we address some limitations of the original method. First, a surface-based regularization is introduced to improve robustness with respect to noise. Second, we design a multi-resolution approach to further improve computational efficiency and to account for large deformations. Third, discrete meshes are used to represent the surface model instead of the implicit level set framework for better computational speed and simpler implementation. Experiment results show that the new morphing active surface model method performs as accurately as a volume registration based re-contouring method but is nearly an order of magnitude faster. The new formulation also allows easy combination of registration and segmentation techniques for further improvement in accuracy and robustness.

We propose a novel approach to fully automatic lesion boundary detection in ultrasound breast images. The novelty of the proposed work lies in the complete automation of the manual process of initial Region-of-Interest (ROI) labeling and in the procedure adopted for the subsequent lesion boundary detection. Histogram equalization is initially used to pre-process the images followed by hybrid filtering and multifractal analysis stages. Subsequently, a single valued thresholding segmentation stage and a rule-based approach is used for the identification of the lesion ROI and the point of interest that is used as the seed-point. Next, starting from this point an Isotropic Gaussian function is applied on the inverted, original ultrasound image. The lesion area is then separated from the background by a thresholding segmentation stage and the initial boundary is detected via edge detection. Finally to further improve and refine the initial boundary, we make use of a state-of-the-art active contour method (i.e. gradient vector flow (GVF) snake model). We provide results that include judgments from expert radiologists on 360 ultrasound images proving that the final boundary detected by the proposed method is highly accurate. We compare the proposed method with two existing state-of- the-art methods, namely the radial gradient index filtering (RGI) technique of Drukker et. al. and the local mean technique proposed by Yap et. al., in proving the proposed method's robustness and accuracy.

Automatic image segmentation techniques are essential for medical image interpretation and analysis. Though numerous methods on image segmentation have been reported, the quality of a segmentation often heavily relies on the positioning of an accurate initial contour. In this paper, a novel solution is presented for the automated object detection in medical image data. A shape- and intensity template is generated from a training set, and both the search image and the template are mapped into a log-polar domain, where rotation and scale are represented by a translation. Orientation and scale of the object are estimated by determining maximum normalized correlation using a Symmetric Phase Only Matched Filter (SPOMF) with a peak enhancement filter. The detected orientation and scale are subsequently applied to the template, and a second pass of the SPOMF using the transformed template yields the actual position of the object in the search image. Performance tests were carried out on two imaging modalities: a set of cardiac MRI images from 34 patients and 2D echocardiograms from 100 patients.

Magnetic Resonance (MR) brain scanning is often planned manually with the goal of aligning the imaging plane with key anatomic landmarks. The planning is time-consuming and subject to inter- and intra- operator variability. An automatic and standardized planning of brain scans is highly useful for clinical applications, and for maximum utility should work on patients of all ages. In this study, we propose a method for fully automatic planning that utilizes the landmarks from two orthogonal images to define the geometry of the third scanning plane. The corpus callosum (CC) is segmented in sagittal images by an active shape model (ASM), and the result is further improved by weighting the boundary movement with confidence scores and incorporating region based refinement. Based on the extracted contour of the CC, several important landmarks are located and then combined with landmarks from the coronal or transverse plane to define the geometry of the third plane. Our automatic method is tested on 54 MR images from 24 patients and 3 healthy volunteers, with ages ranging from 4 months to 70 years old. The average accuracy with respect to two manually labeled points on the CC is 3.54 mm and 4.19 mm, and differed by an average of 2.48 degrees from the orientation of the line connecting them, demonstrating that our method is sufficiently accurate for clinical use.

This paper describes a new system for semi-automatically segmenting the background, subcutaneous fat, interstitial fat, muscle, bone, and bone marrow from magnetic resonance images (MRI's) of volunteers for a new osteoarthritis study. Our system first creates separate right and left thigh images from a single MR image containing both legs. The subcutaneous fat boundary is very difficult to detect in these images and is therefore interactively defined with a single boundary. The volume within the boundary is then automatically processed with a series of clustering and morphological operations designed to identify and classify the different tissue types required for this study. Once the tissues have been identified, the volume of each tissue is determined and a single, false colored, segmented image results. We quantitatively compare the segmentation in three different ways. In our first method we simply compare the tissue volumes of the resulting segmentations performed independently on both the left and right thigh. A second quantification method compares our results temporally with three image sets of the same volunteer made one month apart including a month of leg disuse. Our final quantification methodology compares the volumes of different tissues detected with our system to the results of a manual segmentation performed by a trained expert. The segmented image results of four different volunteers using images acquired at three different times suggests that the system described in this paper provides more consistent results than the manually segmented set. Furthermore, measurements of the left and right thigh and temporal results for both segmentation methods follow the anticipated trend of increasing fat and decreasing muscle over the period of disuse.

We propose Compressed Connected Components (CxCxC), a new fast algorithm for labeling connected components in binary images making use of compression. We break the given 3D image into non-overlapping 2x2x2 cube of voxels (2x2 square of pixels for 2D) and encode these binary values as the bits of a single decimal integer. We perform the connected component labeling on the resulting compressed data set. A recursive labeling approach by the use of smart-masks on the encoded decimal values is performed. The output is finally decompressed back to the original size by decimal-to-binary conversion of the cubes to retrieve the connected components in a lossless fashion. We demonstrate the efficacy of such encoding and labeling for large data sets (up to 1392 x 1040 for 2D and 512 x 512 x 336 for 3D). CxCxC reports a speed gain of 4x for 2D and 12x for 3D with memory savings of 75% for 2D and 88% for 3D over conventional (recursive growing of component labels) connected components algorithm. We also compare our method with those of VTK and ITK and find that we outperform both with speed gains of 3x and 6x for 3D. These features make CxCxC highly suitable for medical imaging and multi-media applications where the size of data sets and the number of connected components can be very large.

The manual segmentation and analysis of 4D high resolution multi slice cardiac CT datasets is both labor intensive and time consuming. Therefore, it is necessary to supply the cardiologist with powerful software tools, to segment the myocardium and the cardiac cavities in all cardiac phases and to compute the relevant diagnostic parameters. In recent years there have been several publications concerning the segmentation and analysis of the left ventricle (LV) and myocardium for a single phase or for the diagnostically most relevant phases, the enddiastole (ED) and the endsystole (ES). However, for a complete diagnosis and especially of wall motion abnormalities, it is necessary to analyze not only the motion endpoints ED and ES, but also all phases in-between. In this paper a novel approach for the 4D segmentation of the left ventricle in cardiac-CT-data is presented. The segmentation of the 4D data is divided into a first part, which segments the motion endpoints of the cardiac cycle ED and ES and a second part, which segments all phases in-between. The first part is based on a bi-temporal statistical shape model of the left ventricle. The second part uses a novel approach based on the individual volume curve for the interpolation between ED and ES and afterwards an active contour algorithm for the final segmentation. The volume curve based interpolation step allows the constraint of the subsequent segmentation of the phases between ED and ES to very small search-intervals, hence makes the segmentation process faster and more robust.

The identification of mammary gland regions is a necessary processing step during the anatomical structure recognition of human body and can be expected to provide the useful information for breast tumor diagnosis. This paper proposes a fully-automated scheme for segmenting the mammary gland regions in non-contrast torso CT images. This scheme calculates the probability for each voxel belonging to the mammary gland or other regions (for example pectoralis major muscles) in CT images and decides the mammary gland regions automatically. The probability is estimated from the location of the mammary gland and pectoralis major muscles in CT images. The location (named as a probabilistic atlas) is investigated from the pre-segmentation results in a number of different CT scans and the CT number distribution is approximated using a Gaussian function. We applied this scheme to 66 patient cases (female, age: 40-80) and evaluated the accuracy by using the coincidence rate between the segmented result and gold standard that is generated manually by a radiologist for each CT case. The mean value of the coincidence rate was 0.82 with the standard deviation of 0.09 for 66 CT cases.

Quantitative analysis of three dimentional (3D) blood flow direction and location will benefit and guide the surgical thinning and dissection process. Toward this goal, this study was performed to reconstruct 3D vascular trees with the incorporation of temporal information from contrast-agent propagation. A computational technique based on our previous work to segment the 3D vascular tree structure from the CT scan volume image sets was proposed. This technique utilizes the deformation method which is a moving grid methodology and which in tradition is used to improve the computational accuracy and efficiency in solving differential equations. Compared with our previous work, we extended the moving grid deformation method to 3D and incorporated 3D region growing method for an initial segmentation. At last, a 3D divergence operator was applied to delineate vascular tree structures from the 3D grid volume plot. Experimental results show the 3D nature of the vascular structure and four-dimensional (4D) vascular tree evolving process. The proposed computational framework demonstrates its effectiveness and improvement in the modeling of 3D vascular tree.

The advancing technology for automatic segmentation of medical images should be accompanied by techniques to inform the user of the local credibility of results. To the extent that this technology produces clinically acceptable segmentations for a significant fraction of cases, there is a risk that the clinician will assume every result is acceptable. In the less frequent case where segmentation fails, we are concerned that unless the user is alerted by the computer, she would still put the result to clinical use. By alerting the user to the location of a likely segmentation failure, we allow her to apply limited validation and editing resources where they are most needed. We propose an automated method to signal suspected non-credible regions of the segmentation, triggered by statistical outliers of the local image match function. We apply this test to m-rep segmentations of the bladder and prostate in CT images using a local image match computed by PCA on regional intensity quantile functions. We validate these results by correlating the non-credible regions with regions that have surface distance greater than 5.5mm to a reference segmentation for the bladder. A 6mm surface distance was used to validate the prostate results. Varying the outlier threshold level produced a receiver operating characteristic with area under the curve of 0.89 for the bladder and 0.92 for the prostate. Based on this preliminary result, our method has been able to predict local segmentation failures and shows potential for validation in an automatic segmentation pipeline.

Many diagnostic problems involve the assessment of vascular structures or bronchial trees depicted in volumetric datasets, but previous algorithms for segmenting cylindrical structures are not sufficiently robust for them to be widely applied clinically. Local geometric information that is of importance in segmentation consists of voxel values and their first and second derivatives. First derivatives can be generalized to the gradient and more generally the structure tensor, while the second derivatives can be represented by Hessian matrices. It is desirable to exploit both kinds of information, at the same time, in any voxel classification process, but few segmentation algorithms have attempted to do this. This project compares segmentation based on the structure tensor to that based on the Hessian matrix, and attempts to determine whether some combination of the two can demonstrate better performance than either individually. To compare performance in a situation where a gold standard exists, the methods were tested on simulated tree structures. We generated 3D tree structures with varying amounts of added noise, and processed them with algorithms based on the structure tensor, the Hessian matrix, and a combination of the two. We applied an orientation-sensitive filter to smooth the tensor fields. The results suggest that the structure tensor by itself is more effective in detecting cylindrical structures than the Hessian tensor, and the combined tensor is better than either of the other tensors.

In most magnetic resonance imaging (MRI) clinical examinations, the orientation and position of diagnostic scans are manually defined by MRI operators. To accelerate the workflow, algorithms have been proposed to automate the definition of the MRI scanning planes. A mid-sagittal plane (MSP), which separates the two cerebral hemispheres, is commonly used to align MRI neurological scans, since it standardizes the visualization of important anatomy. We propose an algorithm to define the MSP automatically based on lines separating the cerebral hemispheres in 2D coronal and transverse images. Challenges to the automatic definition of separation lines are disturbances from the inclusion of the shoulder, and the asymmetry of the brain. The proposed algorithm first detects the position of the head by fitting an ellipse that maximizes the image gradient magnitude in the boundary region of the ellipse. A symmetrical axis is then established which minimizes the difference between the image on either side of the axis. The pixels at the space between the hemispheres are located in the adjacent area of the symmetrical axis, and a linear regression with robust weights defines a line that best separates the two hemispheres. The geometry of MSP is calculated based on the separation lines in the coronal and transverse views. Experiments on 100 images indicate that the result of the proposed algorithm is consistent with the results obtained by domain experts and is significantly faster.

In this paper a study of 3D ultrasound cardiac segmentation using Active Shape Models (ASM) is presented. The proposed approach is based on a combination of a point distribution model constructed from a multitude of high resolution MRI scans and the appearance model obtained from simulated 3D ultrasound images. Usually the appearance model is learnt from a set of landmarked images. The significant level of noise, the low resolution of 3D ultrasound images (3D US) and the frequent failure to capture the complete wall of the left ventricle (LV) makes automatic or manual landmarking difficult. One possible solution is to use artificially simulated 3D US images since the generated images will match exactly the shape in question. In this way, by varying simulation parameters and generating corresponding images, it is possible to obtain a training set where the image matches the shape exactly. In this work the simulation of ultrasound images is performed by a convolutional approach. The evaluation of segmentation accuracy is performed on both simulated and in vivo images. The results obtained on 567 simulated images had an average error of 1.9 mm (1.73 ± 0.05 mm for epicardium and 2 ± 0.07 mm for endocardium, with 95% confidence) with voxel size being 1.1 × 1.1 × 0.7 mm. The error on 20 in vivo data was 3.5 mm (3.44 ± 0.4 mm for epicardium and 3.73 ± 0.4 mm for endocardium). In most images the model was able to approximate the borders of myocardium even when the latter was indistinguishable from the surrounding tissues.

As a result of advances in skin imaging technology and the development of suitable image processing techniques during the last decade, there has been a significant increase of interest in the computer-aided diagnosis of melanoma. Automated border detection is one of the most important steps in this procedure, since the accuracy of the subsequent steps crucially depends on it. In this paper, a fast and unsupervised approach to border detection in dermoscopy images of pigmented skin lesions based on the Statistical Region Merging algorithm is presented. The method is tested on a set of 90 dermoscopy images. The border detection error is quantified by a metric in which a set of dermatologist-determined borders is used as the ground-truth. The proposed method is compared to six state-of-the-art automated methods (optimized histogram thresholding, orientation-sensitive fuzzy c-means, gradient vector flow snakes, dermatologist-like tumor extraction algorithm, meanshift clustering, and the modified JSEG method) and borders determined by a second dermatologist. The results demonstrate that the presented method achieves both fast and accurate border detection in dermoscopy images.

Image interpolation is an essential task in medical imaging because of limited and varying spatial and temporal resolution in imaging devices; also, it is a necessary step while rotating an image for different purposes. In the literature, interpolation techniques have been divided into two major groups: image-based and object-based. Shape-based interpolation is a commonly used object-based method and it works on binary images. In this paper, we propose fuzzy shape-based interpolation by using fuzzy distance transform theory that is applicable to fuzzy object representations. The method essentially works in three steps as follows. Step 1: Separately compute the fuzzy distance transform (FDT) of two successive slices. Step 2: Compute the FDT of the target slice by interpolating the FDT values of original slices. Step 3: Compute the fuzzy object representation on the target slice by applying an inverse FDT (iFDT). Fuzzy shapebased interpolation solves a fundamental problem of shape-based interpolation. Specifically, the new method requires no binarization and it accurately handles the fuzziness of objects using fuzzy distance transform. A new theory and algorithm for iFDT are proposed here to compute the original fuzzy membership map form its fuzzy distance transform map. The idea of iFDT is essential for directly applying the idea of shape-based interpolation to a fuzzy object representation. The method is being tested on clinical data and compared with binary shape-based interpolation.

Many methods are used for warping images to non-rigidly register shapes and objects in between medical images in inter- and intra-patient studies. In landmark-based registration linear methods like thin-plate- or b-splines are often used. These linear methods suffer from a number of theoretical deficiencies: they may break or tear apart the shapes, they are not source-destination symmetric, and may not be invertible. Theoretically more satisfactory models using diffeomorphic approaches like "Large Deformations" and "Brownian warps" have earlier proved (in theory and practice) to remove these deficiencies. In this paper we show that the maximum-likelihood Brownian Warps also generalize better in the case of matching fractured vertebrae to normal vertebrae. X-rays of 10 fractured and 1 normal vertebrae have been annotated by a trained radiologist by 6 so-called height points used for fracture scoring, and by the full boundary. The fractured vertebrae have been registered to the normal vertebra using only the 6 height points as landmarks. After registration the Hausdorff distance between the boundaries is measured. The registrations based on Brownian warps show a significantly lower distance to the original boundary.

Shape analysis has become of increasing interest to the neuroimaging community due to its potential to precisely locate morphological changes between healthy and pathological structures. This manuscript presents a comprehensive set of tools for the computation of 3D structural statistical shape analysis. It has been applied in several studies on brain morphometry, but can potentially be employed in other 3D shape problems. Its main limitations is the necessity of spherical topology. The input of the proposed shape analysis is a set of binary segmentation of a single brain structure, such as the hippocampus or caudate. These segmentations are converted into a corresponding spherical harmonic description (SPHARM), which is then sampled into a triangulated surfaces (SPHARM-PDM). After alignment, differences between groups of surfaces are computed using the Hotelling T² two sample metric. Statistical p-values, both raw and corrected for multiple comparisons, result in significance maps. Additional visualization of the group tests are provided via mean difference magnitude and vector maps, as well as maps of the group covariance information. The correction for multiple comparisons is performed via two separate methods that each have a distinct view of the problem. The first one aims to control the family-wise error rate (FWER) or false-positives via the extrema histogram of non-parametric permutations. The second method controls the false discovery rate and results in a less conservative estimate of the false-negatives. Prior versions of this shape analysis framework have been applied already to clinical studies on hippocampus and lateral ventricle shape in adult schizophrenics. The novelty of this submission is the use of the Hotelling T2 two-sample group difference metric for the computation of a template free statistical shape analysis. Template free group testing allowed this framework to become independent of any template choice, as well as it improved the sensitivity of our method considerably. In addition to our existing correction methodology for the multiple comparison problem using non-parametric permutation tests, we have extended the testing framework to include False Discovery Rate (FDR). FDR provides a significance correction with higher sensitivity while allowing a expected minimal amount of false-positives compared to our prior testing scheme.

We have shown previously that the internal motion caused by a patient's breathing can be estimated from a sequence of slowly rotating 2D cone-beam X-ray projection views and a static prior of of the patient's anatomy.^{1, 2} The estimator iteratively updates a parametric 3D motion model so that the modeled projection views of the deformed reference volume best match the measured projection views. Complicated motion models with many degrees of freedom may better describe the real motion, but the optimizations assiciated with them may overfit noise and may be easily trapped by local minima due to a large number of parameters. For the latter problem, we believe it can be solved by offering the optimization algorithm a good starting point within the valley containing the global minimum point. Therefore, we propose to start the motion estimation with a simplified motion model, in which we assume the displacement of each voxel at any time is proportional to the full movement of that voxel from extreme exhale to extreme inhale. We first obtain the full motion by registering two breath-hold CT volumes at end-expiration and end-inspiration. We then estimate a sequence of scalar displacement proportionality parameters. Thus the goal simplifies to finding a motion amplitude signal. This estimation problem can be solved quickly using the exhale reference volume and projection views with coarse (downsampled) resolution, while still providing acceptable estimation accuracy. The estimated simple motion then can be used to initialize a more complicated motion estimator.

Parallel MRI is a way to use multiple receiver coils with distinct spatial sensitivities to increase the speed of the MRI acquisition. The acquisition is speeded up by undersampling in the phase-encoding direction and the resulting data loss and aliasing is compensated for by the use of the additional information obtained from several receiver coils. The task is to reconstruct an unaliased image from a series of aliased images. We have proposed an algorithm called PROBER that takes advantage of the smoothness of the reconstruction transformation in space. B-spline functions are used to approximate the reconstruction transformation. Their coefficients are estimated at once minimizing the total expected reconstruction error. This makes the reconstruction less sensitive to noise in the reference images and areas without signal in the image. We show that this approach outperforms the SENSE and GRAPPA reconstruction methods for certain coil configurations. In this article, we propose another improvement, consisting of a continuous representation of the B-splines to evaluate the error instead of the discretely sampled version. This solves the undersampling issues in the discrete B-spline representation and offers higher reconstruction quality which has been confirmed by experiments. The method is compared with the discrete version of PROBER and with commercially used algorithms GRAPPA and SENSE in terms of artifact suppression and reconstruction SNR.

Diffusion tensor MR image data gives at each voxel in the image a symmetric, positive definite matrix that is denoted as the diffusion tensor at that voxel location. The eigenvectors of the tensor represent the principal directions of anisotopy in water diffusion. The eigenvector with the largest eigenvalue indicates the local orientation of tissue fibers in 3D as water is expected to diffuse preferentially up and down along the fiber tracts. Although there is no anatomically valid positive or negative direction to these fiber tracts, for many applications, it is of interest to assign an artificial direction to the fiber tract by choosing one of the two signs of the principal eigenvector in such a way that in local neighborhoods the assigned directions are consistent and vary smoothly in space. We demonstrate here an algorithm for realigning the principal eigenvectors by flipping their sign such that it assigns a locally consistent and spatially smooth fiber direction to the eigenvector field based on a Monte-Carlo algorithm adapted from updating clusters of spin systems. We present results that show the success of this algorithm on 11 available unsegmented canine cardiac volumes of both healthy and failing hearts.

The ability to study the biochemical composition of the brain is becoming important to better understand neurodegenerative and neurodevelopmental disorders. Magnetic Resonance Spectroscopy (MRS) can non-invasively provide quantification of brain metabolites in localized regions. The reliability of MRS is limited in part due to partial volume artifacts. This results from the relatively large voxels that are required to acquire sufficient signal-to-noise ratios for the studies. Partial volume artifacts result when a MRS voxel contains a mixture of tissue types. Concentrations of metabolites vary from tissue to tissue. When a voxel contains a heterogeneous tissue composition, the spectroscopic signal acquired from this voxel will consist of the signal from different tissues making reliable measurements difficult. We have developed a novel tool for the estimation of partial volume tissue composition within MRS voxels thus allowing for the correction of partial volume artifacts. In addition, the tool can localize MR spectra to anatomical regions of interest. The tool uses tissue classification information acquired as part of a structural MR scan for the same subject. The tissue classification information is co-registered with the spectroscopic data. The user can quantify the partial volume composition of each voxel and use this information as covariates for metabolite concentrations.

This paper describe our work on tagline detection and tissue strain synthesis. The tagline detection method extends our previous work¹⁶ using pseudo-wavelet reconstruction. The novelty in tagline detection is that we integrated an active contour model and successfully improved the detection and indexing performance. Using pseudo-wavelet reconstruction-based method, prominent wavelet coefficients were retained while others were eliminated. Taglines were then extracted from the reconstructed images using thresholding. Due to noise and artifacts, a tagline can be broken into segments. We employed an active contour model that tracks the most likely segments and bridges them. Experiments demonstrated that our method extracts taglines automatically with greater robustness. Tissue strain was also reconstructed using extracted taglines.

The goal of this project is to use computer analysis to classify small lung nodules, identified on CT, into likely benign and likely malignant categories. We compared discrete wavelet transforms (DWT) based features and a modification of classical features used and reported by others. To determine the best combination of features for classification, several intensities of white noise were added to the original images to determine the effect of such noise on classification accuracy. Two different approaches were used to determine the effect of noise: in the first method the best features for classification of nodules on the original image were retained as noise was added. In the second approach, we recalculated the results to reselect the best classification features for each particular level of added noise. The CT images are from the National Lung Screening Trial (NLST) of the National Cancer Institute (NCI). For this study, nodules were extracted in window frames of three sizes. Malignant nodules were cytologically or histogically diagnosed, while benign had two-year follow-up. A linear discriminant analysis with Fisher criterion (FLDA) approach was used for feature selection and classification, and decision matrix for matched sample to compare the classification accuracy. The initial features mode revealed sensitivity to both the amount of noise and the size of window frame. The recalculated feature mode proved more robust to noise with no change in terms of classification accuracy. This indicates that the best features for computer classification of lung nodules will differ with noise, and, therefore, with exposure.

In this paper, we propose a novel machine learning approach for interactive lesion segmentation on CT and MRI images. Our approach consists of training process and segmenting process. In training process, we train AdaBoosted histogram classifiers to classify true boundary positions and false ones on the 1-D intensity profiles of lesion regions. In segmenting process, given a marker indicating a rough location of a lesion, the proposed solution segments its region automatically by using the trained AdaBoosted histogram classifiers. If there are imperfects in the segmented result, based on one correct location designated by the user, the solution does the segmentation again and gives a new satisfied result. There are two novelties in our approach. The first is that we use AdaBoost in the training process to learn diverse intensity distributions of lesion regions, and utilize the trained classifiers successfully in segmenting process. The second is that we present a reliable and user-friendly way in segmenting process to rectify the segmented result interactively. Dynamic programming is used to find a new optimal path. Experimental results show our approach can segment lesion regions successfully, despite the diverse intensity distributions of the lesion regions, marker location variability and lesion region shape variability. Our framework is also generic and can be applied for blob-like target segmentation with diverse intensity distributions in other applications.

The prevalence of diabetes is expected to increase dramatically in coming years; already today it accounts for a major proportion of the health care budget in many countries. Diabetic Retinopathy (DR), a micro vascular complication very often seen in diabetes patients, is the most common cause of visual loss in working age population of developed countries today. Since the possibility of slowing or even stopping the progress of this disease depends on the early detection of DR, an automatic analysis of fundus images would be of great help to the ophthalmologist due to the small size of the symptoms and the large number of patients. An important symptom for DR are abnormally wide veins leading to an unusually low ratio of the average diameter of arteries to veins (AVR). There are also other diseases like high blood pressure or diseases of the pancreas with one symptom being an abnormal AVR value. To determine it, a classification of vessels as arteries or veins is indispensable. As to our knowledge despite the importance there have only been two approaches to vessel classification yet. Therefore we propose an improved method. We compare two feature extraction methods and two classification methods based on support vector machines and neural networks. Given a hand-segmentation of vessels our approach achieves 95.32% correctly classified vessel pixels. This value decreases by 10% on average, if the result of a segmentation algorithm is used as basis for the classification.

The early detection of Osteoarthritis (OA) treatment efficacy requires monitoring of small changes in cartilage morphology. Current approaches rely in carefully monitoring global cartilage parameters. However, they are not very sensitive to the detection of focal morphological changes in cartilage structure. This work presents the use of the statistical parametric mapping (SPM) for the detection and quantification of changes in cartilage morphology. The SPM is computed by first registering the baseline and the follow-up three dimensional (3D) reconstructions of the cartilage tissue. Once the registration is complete, the thickness changes for every cartilage point is computed which is followed by a model based estimation of the variance of thickness error. The cartilage thickness change and the variance estimations are used to compute the z-score map. The map is used to visualize and quantify significant changes in cartilage thickness. The z-map quantification provides the area of significant changes, the associated volume of changes as well as the average thickness of cartilage loss. Furthermore, thickness change distributions functions are normalized to provide the probability distribution functions (PDF). The PDF can be used to understand and quantify the differences among different treatment groups. The performance of the approach on simulated data and real subject data will be presented.

The performance of classification algorithms for differentiating among obstructive lung diseases based on features from texture analysis using HRCT (High Resolution Computerized Tomography) images was compared. HRCT can provide accurate information for the detection of various obstructive lung diseases, including centrilobular emphysema, panlobular emphysema and bronchiolitis obliterans. Features on HRCT images can be subtle, however, particularly in the early stages of disease, and image-based diagnosis is subject to inter-observer variation. To automate the diagnosis and improve the accuracy, we compared three types of automated classification systems, naïve Bayesian classifier, ANN (Artificial Neural Net) and SVM (Support Vector Machine), based on their ability to differentiate among normal lung and three types of obstructive lung diseases. To assess the performance and cross-validation of these three classifiers, 5 folding methods with 5 randomly chosen groups were used. For a more robust result, each validation was repeated 100 times. SVM showed the best performance, with 86.5% overall sensitivity, significantly different from the other classifiers (one way ANOVA, p<0.01). We address the characteristics of each classifier affecting performance and the issue of which classifier is the most suitable for clinical applications, and propose an appropriate method to choose the best classifier and determine its optimal parameters for optimal disease discrimination. These results can be applied to classifiers for differentiation of other diseases.

Texture analysis and classification of soft tissues in Computed Tomography (CT) images recently advanced with a new approach that disambiguates the checkboard problem where two distinctly different patterns produce identical co-occurrence matrices, but this method quadruples the size of the feature space. The feature space size problem is exacerbated by the use of varying sized texture operators for improving boundary segmentation. Dimensionality reduction motivates this investigation into systematic analysis of the power of feature categories (Haralick descriptors, distance, and direction) to differentiate between soft tissues. The within-organ variance explained by the individual components of feature categories offers a ranking of their potential power for between-organ discrimination. This paper introduces a technique for combining the Principal Component Analysis (PCA) results to compare and visualize the explanatory power of features with varying window sizes. We found that 1) the two Haralick features Cluster Tendency and Contrast contribute the most; 2) as distance increases, its contribution to overall variance decreases; and 3) direction is unimportant. We also evaluated the proposed technique with respect to its classification power. Linear Discriminant Analysis (LDA) and Decision Tree (DT) were used to produce two classification models based on the reduced data set. We found that using PCA either fails to improve or markedly degrades the classification performance of LDA as well as of the DT model. Though feature extraction for classification shows no promise, the proposed technique offers a systematic mechanism to compare feature reduction strategies for varying window sizes as well as other measurement techniques.

The visualisation and subsequent quantification of the inner bone structure plays an important role for better understanding the disease- or drug-induced changes of the bone in the context of osteoporosis. Scaling indices (SIM) are well suited to quantify these structures on a local level, especially to discriminate between plate-like and rod-like structural elements. Local filters based on wavelets (WVL) are a standard technique in texture analysis. So far, they are mainly used for two-dimensional image data sets. Here we extend the formalism of the spherical Mexican hat wavelets to the analysis of three-dimensional tomographic images and evaluate its performance in comparison with scaling indices, histomorphometric measures and BMD. &mgr;CT images with isotropic resolution of 30 x 30 x 30 &mgr;m of a sample of 19 trabecular bone specimen of human thoracic vertebrae were acquired. In addition, the bone mineral density was measured by QCT. The maximum compressive strength (MCS) was determined in a biomechanical test. Some wavelet-based as well as all scaling index- based texture measures show a significantly higher correlation with MCS (WVL: &rgr;²=0.54, SIM: &rgr;²=0.53-0.56) than BMD (&rgr;²=0.46), where we find slightly better correlations for SIM than for WVL. The SIM and WVL results are comparable but not better to those obtained with histomorphometric measures (BV/TV: &rgr;²=0.45, Tr. N.: &rgr;²=0.67, Tr.Sp.: &rgr;²=0.67). In conclusion, WVL and SIM techniques can successfully be applied to &mgr;CT image data. Since the two measures characterize the image structures on a local scale, they offer the possibility to directly identify and discriminate rods and sheets of the trabecular structure. This property may give new insights about the bone constituents responsible for the mechanical strength.

In this paper we focus on the problem of automatic classification of melanocytic lesions, aiming at identifying the presence of reticular patterns. The recognition of reticular lesions is an important step in the description of the pigmented network, in order to obtain meaningful diagnostic information. Parameters like color, size or symmetry could benefit from the knowledge of having a reticular or non-reticular lesion. The detection of network patterns is performed with a three-steps procedure. The first step is the localization of line points, by means of the line points detection algorithm, firstly described by Steger. The second step is the linking of such points into a line considering the direction of the line at its endpoints and the number of line points connected to these. Finally a third step discards the meshes which couldn't be closed at the end of the linking procedure and the ones characterized by anomalous values of area or circularity. The number of the valid meshes left and their area with respect to the whole area of the lesion are the inputs of a discriminant function which classifies the lesions into reticular and non-reticular. This approach was tested on two balanced (both sets are formed by 50 reticular and 50 non-reticular images) training and testing sets. We obtained above 86% correct classification of the reticular and non-reticular lesions on real skin images, with a specificity value never lower than 92%.

Fragility fractures or pathologic fractures of the hip, i.e. fractures with no apparent trauma, represent the worst complication in osteoporosis with a mortality close to 25% during the first post-traumatic year. Over 90% of hip fractures result from falls from standing height. A substantial number of femoral fractures are initiated in the femoral neck or the trochanteric regions which contain an internal architecture of trabeculae that are functionally highly specialized to withstand the complex pattern of external and internal forces associated with human gait. Prediction of the mechanical strength of bone tissue can be achieved by dedicated texture analysis of data obtained by high resolution imaging modalities, e.g. computed tomography (CT) or magnetic resonance tomography (MRI). Since in the case of the proximal femur, the connectivity, regional distribution and - most of all - the preferred orientation of individual trabeculae change considerably within narrow spatial limits, it seems most reasonable to evaluate the femoral bone structure on an orientation-weighted, local scale. In past studies, we could demonstrate the advantages of topological analysis of bone structure using the Minkowski Functionals in 3D on a global and on a local scale. The current study was designed to test the hypothesis that the prediction of the mechanical competence of the proximal femur by a new algorithm considering orientational changes of topological properties in the trabecular architecture is feasible and better suited than conventional methods based on the measurement of the mineral density of bone tissue (BMD).

Automatic liver segmentation from abdominal computed tomography (CT) images based on gray levels or shape alone is difficult because of the overlap in gray-level ranges and the variation in position and shape of the soft tissues. To address these issues, we propose an automatic liver segmentation method that utilizes low-level features based on texture information; this texture information is expected to be homogenous and consistent across multiple slices for the same organ. Our proposed approach consists of the following steps: first, we perform pixel-level texture extraction; second, we generate liver probability images using a binary classification approach; third, we apply a split-and-merge algorithm to detect the seed set with the highest probability area; and fourth, we apply to the seed set a region growing algorithm iteratively to refine the liver's boundary and get the final segmentation results. Furthermore, we compare the segmentation results from three different texture extraction methods (Co-occurrence Matrices, Gabor filters, and Markov Random Fields (MRF)) to find the texture method that generates the best liver segmentation. From our experimental results, we found that the co-occurrence model led to the best segmentation, while the Gabor model led to the worst liver segmentation. Moreover, co-occurrence texture features alone produced approximately the same segmentation results as those produced when all the texture features from the combined co-occurrence, Gabor, and MRF models were used. Therefore, in addition to providing an automatic model for liver segmentation, we also conclude that Haralick cooccurrence texture features are the most significant texture characteristics in distinguishing the liver tissue in CT scans.

In this paper, we proposed novel shape features to improve classification performance of differentiating obstructive lung diseases, based on HRCT (High Resolution Computerized Tomography) images. The images were selected from HRCT images, obtained from 82 subjects. For each image, two experienced radiologists selected rectangular ROIs with various sizes (16x16, 32x32, and 64x64 pixels), representing each disease or normal lung parenchyma. Besides thirteen textural features, we employed additional seven shape features; cluster shape features, and Top-hat transform features. To evaluate the contribution of shape features for differentiation of obstructive lung diseases, several experiments were conducted with two different types of classifiers and various ROI sizes. For automated classification, the Bayesian classifier and support vector machine (SVM) were implemented. To assess the performance and cross-validation of the system, 5-folding method was used. In comparison to employing only textural features, adding shape features yields significant enhancement of overall sensitivity(5.9, 5.4, 4.4% in the Bayesian and 9.0, 7.3, 5.3% in the SVM), in the order of ROI size 16x16, 32x32, 64x64 pixels, respectively (t-test, p<0.01). Moreover, this enhancement was largely due to the improvement on class-specific sensitivity of mild centrilobular emphysema and bronchiolitis obliterans which are most hard to differentiate for radiologists. According to these experimental results, adding shape features to conventional texture features is much useful to improve classification performance of obstructive lung diseases in both Bayesian and SVM classifiers.

We applied a new texture segmentation algorithm to improve the segmentation of boundary areas for distinction on the liver needle biopsy images taken from microscopes for automatic assessment of liver fibrosis severity. It was difficult to gain satisfactory segmentation results on the boundary areas of textures with some of existing texture segmentation algorithms in our preliminary experiments. The proposed algorithm consists of three steps. The first step is to apply the K-View-datagram segmentation method to the image. The second step is to find a boundary set which is defined as a set including all the pixels with more than half of its neighboring pixels being classified into clusters other than that of itself by the K-View-datagram method. The third step is to apply a modified K-view template method with a small scanning window to the boundary set to refine the segmentation. The algorithm was applied to the real liver needle biopsy images provided by the hospitals in Wuhan, China. Initial experimental results show that this new segmentation algorithm gives high segmentation accuracy and classifies the boundary areas better than the existing algorithms. It is a useful tool for automatic assessment of liver fibrosis severity.

Osteoporosis is a metabolic bone disorder characterized by the loss of bone mineral density (BMD) and the deterioration of the bone micro-architecture. Rarefied bone structures are more susceptible to fractures which are the worst complications of osteoporosis. Here, we apply a structure characterization method, namely the Scaling Index Method, to micro-computed tomographic (&mgr;-CT) images of the distal radius and extract 3D nonlinear structure measures to assess the biomechanical properties of trabecular bone. Biomechanical properties were quantified by the maximum compressive strength (MCS) obtained in a biomechanical test and bone mineral density (BMD) was calculated using dual X-ray absorptiometry (DXA). &mgr;-CT images allow for the application of two different modalities of the SIM which differ in the dimensional embedding of the image. Both representations lead to similar correlation coefficients with MCS which are significantly better than the ones obtained using standard 3D morphometric parameters and comparable to the result given by BMD. The analysis of &mgr;-CT images based on the SIM allows for a sharp distinction of the different structural elements which compose the trabecular bone network.

Cervicography is a technique for visual screening of uterine cervix images for cervical cancer. One of our research goals is the automated detection in these images of acetowhite (AW) lesions, which are sometimes correlated with cervical cancer. These lesions are characterized by the whitening of regions along the squamocolumnar junction on the cervix when treated with 5% acetic acid. Image preprocessing is required prior to invoking AW detection algorithms on cervicographic images for two reasons: (1) to remove Specular Reflections (SR) caused by camera flash, and (2) to isolate the cervix region-of-interest (ROI) from image regions that are irrelevant to the analysis. These image regions may contain medical instruments, film markup, or other non-cervix anatomy or regions, such as vaginal walls. We have qualitatively and quantitatively evaluated the performance of alternative preprocessing algorithms on a test set of 120 images. For cervix ROI detection, all approaches use a common feature set, but with varying combinations of feature weights, normalization, and clustering methods. For SR detection, while one approach uses a Gaussian Mixture Model on an intensity/saturation feature set, a second approach uses Otsu thresholding on a top-hat transformed input image. Empirical results are analyzed to derive conclusions on the performance of each approach.

Small errors in the alignment between CT Attenuation Correction (CTAC) images and Positron Emission Tomography (PET) acquisitions can result in significant changes in PET attenuation corrected images. Misalignment due to respiratory or cardiac motion can produce mismatch between the PET and CTAC acquisitions. This contributes to artifactual hypoperfusion defects that are interpretable as myocardial ischemia or infarct. Correction for the misalignment between the PET and CTAC images can eliminate these false positive artifacts. Two methods for correcting for this respiratory and cardiac misalignment were compared. The first was an existing procedure, the manual-shift method, using point-to-point, in-plane, two-dimensional (2D) measurements of the shifts in axial, sagittal, and coronal planes. A new PET image reconstruction using the corrected attenuation map shifted by the 2D measurements was then performed. In the second method, the Interactive ACQC method, visual alignment was performed between the left ventricle boundaries on fused images and automated calculation of necessary rigid three-dimensional (3D) alignment parameters was performed. A new PET image reconstruction was then performed with an attenuation map shifted by the prescribed alignment parameters. The two methods were compared for accuracy and workflow efficiency using five cardiac PET/CT cases, scanned on GE Discovery VCT and Discovery ST systems. Alignment measurements using the visual alignment process (the interactive ACQC method) improved productivity by over five minutes, on average. The results show that the interactive ACQC procedure yields similar results to those of the point-to-point procedure while providing improved workflow for cardiac PET attenuation correction quality control.

The JPEG2000 standard is currently widely adopted in medical and volumetric data compression. In this respect, a 3D extension (JPEG2000 Part 10 - JP3D) is currently being standardized. However, no suitable 3D context model is yet available within the standard, such that the context-based arithmetic entropy coder of JP3D still uses the 2D context model of JPEG2000 Part 1. In this paper, we propose a context design algorithm that, based on a training set, generates an optimized 3D context model, while avoiding an exhaustive search and at the same time keeping the space and time complexities well within the limits of today hardware. The algorithm comes as a solution for the situations in which the number of allowable initial contexts is very large. In this sense, the three-dimensional 3x3x3 context neighborhood investigated in this paper is a good example of an instantiation that would have otherwise been computationally unfeasible. Furthermore, we have designed a new 3D context model for JP3D. We show that the JP3D codec equipped with this model consistently outperforms its 2D context model counterpart, for an extended test dataset. In this respect, we report a gain in lossless compression performance of up to 10%. Moreover, for a large range of bitrates, we always obtain gains in PSNR, sometimes even over 3dB.

The size of medical data has increased significantly over the last few years. This poses severe problems for the rapid transmission of medical data across the hospital network resulting into longer access times of the images. Also longterm storage of data becomes more and more a problem. In an attempt to overcome the increasing data size often lossless or lossy compression algorithms are being used. This paper compares the existing JPEG2000 compression algorithm and the new emerging JP3D standard for compression of volumetric datasets. The main benefit of JP3D is that this algorithm truly is a 3D compression algorithm that exploits correlation not only within but also in between slices of a dataset. We evaluate both lossless and lossy modes of these algorithms. As a first step we perform an objective evaluation. Using RMSE and PSNR metrics we determine which compression algorithm performs best and this for multiple compression ratios and for several clinically relevant medical datasets. It is well known that RMSE and PSNR often do not correlate well with subjectively perceived image quality. Therefore we also perform a psycho visual analysis by means of a numerical observer. With this observer model we analyze how compression artifacts actually are perceived by a human observer. Results show superior performance of the new JP3D algorithm compared to the existing JPEG2000 algorithm.

Medical imaging poses the great challenge of having compression algorithms that are lossless for diagnostic and legal reasons and yet provide high compression rates for reduced storage and transmission time. The images usually consist of a region of interest representing the part of the body under investigation surrounded by a "background", which is often noisy and not of diagnostic interest. In this paper, we propose a ROI-based 3D coding system integrating both the segmentation and the compression tools. The ROI is extracted by an atlas based 3D segmentation method combining active contours with information theoretic principles, and the resulting segmentation map is exploited for ROI based coding. The system is equipped with a GUI allowing the medical doctors to supervise the segmentation process and eventually reshape the detected contours at any point. The process is initiated by the user through the selection of either one pre-de.ned reference image or one image of the volume to be used as the 2D "atlas". The object contour is successively propagated from one frame to the next where it is used as the initial border estimation. In this way, the entire volume is segmented based on a unique 2D atlas. The resulting 3D segmentation map is exploited for adaptive coding of the different image regions. Two coding systems were considered: the JPEG3D standard and the 3D-SPITH. The evaluation of the performance with respect to both segmentation and coding proved the high potential of the proposed system in providing an integrated, low-cost and computationally effective solution for CAD and PAC systems.

Two common approaches have been developed to compress volumetric medical data from sources such as magnetic resonance imaging (MRI) and computed tomography (CT): (1) 2D-based compression methods, which compress each image slice independently using 2D image codecs; and (2) 3D-based compression methods, which treat the data as true volumetric data and compress using 3D image codecs. It has been shown that most 3D-based compression methods, such as 3D-SPIHT, can achieve significantly higher compression quality than most 2D-based compression methods, such as JPEG, JPEG-2000, and 2D-SPIHT. However, the compression/decompression speed is slow, and the high computational complexity and high memory usage render 3D-based compressions difficult to implement in hardware. In this paper, we propose a new 3D-based compression algorithm, 3D-BCWT, which is an extension to the computationally efficient BCWT (Backward Coding of Wavelet Trees) algorithm [10]. 3D-BCWT not only can achieve the same high compression quality as 3D-SPIHT does, but it can also provide extremely fast compression/decompression speed, low complexity, and low memory usage, which are ideal for low-cost hardware and software implementations and for compressing high resolution volumetric data. Moreover, 3D-BCWT also possesses the capabilities of progressive transmission and decoding, such as progression of resolution and progression of quality, which are essential features for efficient image retrieval from large online archives.

In this paper, we propose a compression scheme that is tailored for stereo-laparoscope sequences. The inter-frame correlation is modeled by the deformation field obtained by elastic registration between two subsequent frames and exploited for prediction of the left sequence. The right sequence is lossy encoded by prediction from the corresponding left images. Wavelet-based coding is applied to both the deformation vector fields and residual images. The resulting system supports spatio temporal scalability, while providing lossless performance. The implementation of the wavelet transform by integer lifting ensures a low computational complexity, thus reducing the required run-time memory allocation and on line implementation. Extensive psychovisual tests were performed for system validation and characterization with respect to the MPEG4 standard for video coding. Results are very encouraging: the PSVC system features the functionalities making it suitable for PACS while providing a good trade-off between usability and performance in lossy mode.

Breast tomosynthesis is an emerging state-of-the-art three-dimensional (3D) imaging technology that demonstrates significant early promise in screening and diagnosing breast cancer. However, this kind of image has significant out-of-plane artifacts due to its limited tomography nature, which affects the image quality and further would interrupt interpretation. In this paper, we develop a robust deblurring method to remove or suppress blurry artifacts by applying three-dimensional (3D) nonlinear anisotropic diffusion filtering method. Differential equation of 3D anisotropic diffusion filtering is discretized using explicit and implicit numerical methods, respectively, combined by first (fixed grey value) and second (adiabatic) boundary conditions under ten nearest neighbor grids configuration of finite difference scheme. The discretized diffusion equation is applied in the breast volume reconstructed from the entire tomosynthetic images of breast. The proposed diffusion filtering method is evaluated qualitatively and quantitatively on clinical tomosynthesis images. Results indicate that the proposed diffusion filtering method is very powerful in suppressing the blurry artifacts, and the results also indicate that implicit numerical algorithm with fixed value boundary condition has better performance in enhancing the contrast of tomosynthesis image, demonstrating the effectiveness of the proposed filtering method in deblurring the out-of-plane artifacts.

We are developing a computerized technique to reduce intra- and interplane ghosting artifacts caused by high-contrast objects such as dense microcalcifications (MCs) or metal markers on the reconstructed slices of digital tomosynthesis mammography (DTM). In this study, we designed a constrained iterative artifact reduction method based on a priori 3D information of individual MCs. We first segmented individual MCs on projection views (PVs) using an automated MC detection system. The centroid and the contrast profile of the individual MCs in the 3D breast volume were estimated from the backprojection of the segmented individual MCs on high-resolution (0.1 mm isotropic voxel size) reconstructed DTM slices. An isolated volume of interest (VOI) containing one or a few MCs is then modeled as a high-contrast object embedded in a local homogeneous background. A shift-variant 3D impulse response matrix (IRM) of the projection-reconstruction (PR) system for the extracted VOI was calculated using the DTM geometry and the reconstruction algorithm. The PR system for this VOI is characterized by a system of linear equations. A constrained iterative method was used to solve these equations for the effective linear attenuation coefficients (eLACs) within the isolated VOI. Spatial constraint and positivity constraint were used in this method. Finally, the intra- and interplane artifacts on the whole breast volume resulting from the MC were calculated using the corresponding impulse responses and subsequently subtracted from the original reconstructed slices. The performance of our artifact-reduction method was evaluated using a computer-simulated MC phantom, as well as phantom images and patient DTMs obtained with IRB approval. A GE prototype DTM system that acquires 21 PVs in 3º increments over a ±30º range was used for image acquisition in this study. For the computer-simulated MC phantom, the eLACs can be estimated accurately, thus the interplane artifacts were effectively removed. For MCs in phantom and patient DTMs, our method reduced the artifacts but also created small over-corrected areas in some cases. Potential reasons for this may include: the simplified mathematical modeling of the forward projection process, and the amplified noise in the solution of the system of linear equations.

Temporal subtraction is a visual enhancement technique to improve the detection of pathological changes from medical images acquired at different times. Prior to subtracting a previous image from a current image, a nonrigid warping of the two images might be necessary. As the nonrigid warping may change the size of pathological lesions, the subtraction image can be misleading. In this paper we present an alternative subtraction technique to avoid this problem. Instead of subtracting the intensities of corresponding voxels, a convolution filter is applied to both images prior to subtraction. The technique is demonstrated for computed tomography images of the lungs. It is shown that this method results in an improved visual enhancement of changing nodules compared with the conventional subtraction technique.