This PDF file contains the front matter associated with SPIE Proceedings Volume 10134, including the Title Page, Copyright information, Table of Contents, and Conference Committee listing.

We are developing a computerized system for detection of bladder cancer in CT urography. In this study, we used a deep-learning convolutional neural network (DL-CNN) to segment the bladder wall. This task is challenging due to differences in the wall between the contrast and non-contrast-filled regions, significant variations in appearance, size, and shape of the bladder among cases, overlap of the prostate with the bladder wall, and the wall being extremely thin compared to the overall size of the bladder. We trained a DL-CNN to estimate the likelihood that a given pixel would be inside the wall of the bladder using neighborhood information. A segmented bladder wall was then obtained using level sets with this likelihood map as a term in the level set energy formulation to obtain contours of the inner and outer bladder walls. The accuracy of the segmentation was evaluated by comparing the segmented wall outlines to hand outlines for a set of 79 training cases and 15 test cases using the average volume intersection % as the metric. For the training set, the inner wall achieved an average volume intersection of 90.0±8.7% and the outer wall achieved 93.7±3.9%. For the test set, the inner wall achieved an average volume intersection of 87.6±7.6% and the outer wall achieved 87.2±9.3%. The results show that the DL-CNN with level sets was effective in segmenting the inner and outer bladder walls.

We are developing a computer-aided detection system for bladder cancer in CT urography (CTU). We have previously developed methods for detection of bladder masses within the contrast-enhanced and the non-contrastenhanced regions of the bladder individually. In this study, we investigated methods for detection of bladder masses within the entire bladder. The bladder was segmented using our method that combined deep-learning convolutional neural network with level sets. The non-contrast-enhanced region was separated from the contrast-enhanced region with a maximum-intensity-projection-based method. The non-contrast region was smoothed and gray level threshold was applied to the contrast and non-contrast regions separately to extract the bladder wall and potential masses. The bladder wall was transformed into a straightened thickness profile, which was analyzed to identify lesion candidates in a prescreening step. The candidates were mapped back to the 3D CT volume and segmented using our auto-initialized cascaded level set (AI-CALS) segmentation method. Twenty-seven morphological features were extracted for each candidate. A data set of 57 patients with 71 biopsy-proven bladder lesions was used, which was split into independent training and test sets: 42 training cases with 52 lesions, and 15 test cases with 19 lesions. Using the training set, feature selection was performed and a linear discriminant (LDA) classifier was designed to merge the selected features for classification of bladder lesions and false positives. The trained classifier was evaluated with the test set. FROC analysis showed that the system achieved a sensitivity of 86.5% at 3.3 FPs/case for the training set, and 84.2% at 3.7 FPs/case for the test set.

We are developing a CAD system for bladder cancer treatment response assessment in CT. We compared the performance of the deep-learning convolution neural network (DL-CNN) using different network sizes, and with and without transfer learning using natural scene images or regions of interest (ROIs) inside and outside the bladder. The DL-CNN was trained to identify responders (T0 disease) and non-responders to chemotherapy. ROIs were extracted from segmented lesions in pre- and post-treatment scans of a patient and paired to generate hybrid pre-post-treatment paired ROIs. The 87 lesions from 82 patients generated 104 temporal lesion pairs and 6,700 pre-post-treatment paired ROIs. Two-fold cross-validation and receiver operating characteristic analysis were performed and the area under the curve (AUC) was calculated for the DL-CNN estimates. The AUCs for prediction of T0 disease after treatment were 0.77±0.08 and 0.75±0.08, respectively, for the two partitions using DL-CNN without transfer learning and a small network, and were 0.74±0.07 and 0.74±0.08 with a large network. The AUCs were 0.73±0.08 and 0.62±0.08 with transfer learning using a small network pre-trained with bladder ROIs. The AUC values were 0.77±0.08 and 0.73±0.07 using the large network pre-trained with the same bladder ROIs. With transfer learning using the large network pretrained with the Canadian Institute for Advanced Research (CIFAR-10) data set, the AUCs were 0.72±0.06 and 0.64±0.09, respectively, for the two partitions. None of the differences in the methods reached statistical significance. Our study demonstrated the feasibility of using DL-CNN for the estimation of treatment response in CT. Transfer learning did not improve the treatment response estimation. The DL-CNN performed better when transfer learning with bladder images was used instead of natural scene images.

Prostate cancer (PCa) is the second most common cause of cancer related deaths in men. Multiparametric MRI (mpMRI) is the most accurate imaging method for PCa detection; however, it requires the expertise of experienced radiologists leading to inconsistency across readers of varying experience. To increase inter-reader agreement and sensitivity, we developed a computer-aided detection (CAD) system that can automatically detect lesions on mpMRI that readers can use as a reference. We investigated a convolutional neural network based deep-learing (DCNN) architecture to find an improved solution for PCa detection on mpMRI. We adopted a network architecture from a state-of-the-art edge detector that takes an image as an input and produces an image probability map. Two-fold cross validation along with a receiver operating characteristic (ROC) analysis and free-response ROC (FROC) were used to determine our deep-learning based prostate-CAD’s (CADDL) performance. The efficacy was compared to an existing prostate CAD system that is based on hand-crafted features, which was evaluated on the same test-set. CADDL had an 86% detection rate at 20% false-positive rate while the top-down learning CAD had 80% detection rate at the same false-positive rate, which translated to 94% and 85% detection rate at 10 false-positives per patient on the FROC. A CNN based CAD is able to detect cancerous lesions on mpMRI of the prostate with results comparable to an existing prostate-CAD showing potential for further development.

Functional analysis of the placenta is important to analyze and understand its role in fetal growth and development. BOLD MR is a non-invasive technique that has been extensively used for functional analysis of the brain. During the last years, several studies have shown that this dynamic image modality is also useful to extract functional information of the placenta. We propose in this paper a method to track the placenta from a sequence of BOLD MR images acquired under normoxia and hyperoxia conditions with the goal of quantifying how the placenta adapts to oxygenation changes. The method is based on a spatiotemporal transformation model that ensures temporal coherence of the tracked structures. The method was initially applied to four patients with healthy pregnancies. An average MR signal increase of 16.96±8.39% during hyperoxia was observed. These automated results are in line with state-of-the-art reports using time-consuming manual segmentations subject to inter-observer errors.

The infiltrative lung diseases are a class of irreversible, non-neoplastic lung pathologies requiring regular follow-up with CT imaging. Quantifying the evolution of the patient status imposes the development of automated classification tools for lung texture. Traditionally, such classification relies on a two-dimensional analysis of axial CT images. This paper proposes a cascade of the existing CNN based CAD system, specifically tuned-up. The advantage of using a deep learning approach is a better regularization of the classification output. In a preliminary evaluation, the combined approach was tested on a 13 patient database of various lung pathologies, showing an increase of 10% in True Positive Rate (TPR) with respect to the best suited state of the art CNN for this task.

A fully-automated computer algorithm has been developed to identify early-stage Usual Interstitial Pneumonia (UIP) using features computed from low-dose CT scans. In each scan, the pre-segmented lung region is divided into N subsections (N = 1, 8, 27, 64) by separating the lung from anterior/posterior, left/right and superior/inferior in 3D space. Each subsection has approximately the same volume. In each subsection, a classic density measurement (fractional high-density volume h) is evaluated to characterize the disease severity in that subsection, resulting in a feature vector of length N for each lung. Features are then combined in two different ways: concatenation (2*N features) and taking the maximum in each of the two corresponding subsections in the two lungs (N features).

The algorithm was evaluated on a dataset consisting of 51 UIP and 56 normal cases, a combined feature vector was computed for each case and an SVM classifier (RBF kernel) was used to classify them into UIP or normal using ten-fold cross validation. A receiver operating characteristic (ROC) area under the curve (AUC) was used for evaluation. The highest AUC of 0.95 was achieved by using concatenated features and an N of 27. Using lung partition (N = 27, 64) with concatenated features had significantly better result over not using partitions (N = 1) (p-value < 0.05). Therefore this equal-volume partition fractional high-density volume method is useful in distinguishing early-stage UIP from normal cases.

Deep convolutional neural networks (CNNs) form the backbone of many state-of-the-art computer vision systems for classification and segmentation of 2D images. The same principles and architectures can be extended to three dimensions to obtain 3D CNNs that are suitable for volumetric data such as CT scans. In this work, we train a 3D CNN for automatic detection of pulmonary nodules in chest CT images using volumes of interest extracted from the LIDC dataset. We then convert the 3D CNN which has a fixed field of view to a 3D fully convolutional network (FCN) which can generate the score map for the entire volume efficiently in a single pass. Compared to the sliding window approach for applying a CNN across the entire input volume, the FCN leads to a nearly 800-fold speed-up, and thereby fast generation of output scores for a single case. This screening FCN is used to generate difficult negative examples that are used to train a new discriminant CNN. The overall system consists of the screening FCN for fast generation of candidate regions of interest, followed by the discrimination CNN.

Lung cancer is the leading cause of cancer deaths worldwide. Early diagnosis is critical in increasing the 5-year survival rate of lung cancer, so the efficient and accurate detection of lung nodules, potential precursors to lung cancer, is evermore important. In this paper, a computer-aided lung nodule detection system using convolution neural networks (CNN) and handcrafted features for false positive reduction is developed. The CNNs were trained with three types of images: lung CT images, their nodule-enhanced images, and their blood vessel-enhanced images. For each nodule candidate, nine 2D patches from differently oriented planes were extracted from each type of images. Patches of the same orientation from the same type of image across different candidates were used to train the CNNs independently, which were used to extract 864 features. 88 handcrafted features including intensity, shape, and texture features were also obtained from the lung CT images. The CNN features and handcrafted features were then combined to train a classifier, and a support vector machine was adopted to achieve the final classification results. The proposed method was evaluated on 1004 CT scans from the LIDC-IDRI database using 10-fold cross-validation. Compared with the traditional CNN method using only lung CT images, the proposed method boosted the sensitivity of nodule detection from 89.0% to 90.9% at 4 FPs/scan and from 71.6% to 78.2% at 1 FP/scan. This indicates that a combination of handcrafted features and CNN features from both lung CT images and enhanced images is a promising method for lung nodule detection.

For pediatric oncology patients, CT scans are performed to assess treatment response and disease progression. CAD may be used to detect lung nodules which would reflect metastatic disease. The purpose of this study was to investigate the effects of reducing radiation dose and varying slice thickness on CAD performance in the detection of solid lung nodules in pediatric patients. The dataset consisted of CT scans of 58 pediatric chest cases, from which 7 cases had lung nodules detected by radiologist, and a total of 28 nodules were marked. For each case, the original raw data (sinogram data) was collected and a noise addition model was used to simulate reduced-dose scans of 50%, 25% and 10% of the original dose. In addition, the original and reduced-dose raw data were reconstructed at slice thicknesses of 1.5 and 3 mm using a medium sharp (B45) kernel; the result was eight datasets (4 dose levels x 2 thicknesses) for each case An in-house CAD tool was applied on all reconstructed scans, and results were compared with the radiologist’s markings. Patient level mean sensitivities at 3mm thickness were 24%, 26%, 25%, 27%, and at 1.5 mm thickness were 23%, 29%, 35%, 36% for 10%, 25%, 50%, and 100% dose level, respectively. Mean FP numbers were 1.5, 0.9, 0.8, 0.7 at 3 mm and 11.4, 3.5, 2.8, 2.8 at 1.5 mm thickness for 10%, 25%, 50%, and 100% dose level respectively. CAD sensitivity did not change with dose level for 3mm thickness, but did change with dose for 1.5 mm. False Positives increased at low dose levels where noise values were high.

In this study, we propose a novel method of lung lesion detection in FDG-PET/CT volumes without labeling lesions. In our method, the probability distribution over normal standardized uptake values (SUVs) is estimated from the features extracted from the corresponding volume of interest (VOI) in the CT volume, which include gradient-based and texture-based features. To estimate the distribution, we use Gaussian process regression with an automatic relevance determination kernel, which provides the relevance of feature values to estimation. Our model was trained using FDG-PET/CT volumes of 121 normal cases. In the lesion detection phase, the actual SUV is judged as normal or abnormal by comparison with the estimated SUV distribution. According to the validation using 28 FDG-PET/CT volumes with 34 lung lesions, the sensitivity of the proposed method at 5.0 false positives per case was 81.9%.

Volumetric Laser Endomicroscopy (VLE) is a promising technique for the detection of early neoplasia in Barrett’s Esophagus (BE). VLE generates hundreds of high resolution, grayscale, cross-sectional images of the esophagus. However, at present, classifying these images is a time consuming and cumbersome effort performed by an expert using a clinical prediction model. This paper explores the feasibility of using computer vision techniques to accurately predict the presence of dysplastic tissue in VLE BE images. Our contribution is threefold. First, a benchmarking is performed for widely applied machine learning techniques and feature extraction methods. Second, three new features based on the clinical detection model are proposed, having superior classification accuracy and speed, compared to earlier work. Third, we evaluate automated parameter tuning by applying simple grid search and feature selection methods. The results are evaluated on a clinically validated dataset of 30 dysplastic and 30 non-dysplastic VLE images. Optimal classification accuracy is obtained by applying a support vector machine and using our modified Haralick features and optimal image cropping, obtaining an area under the receiver operating characteristic of 0.95 compared to the clinical prediction model at 0.81. Optimal execution time is achieved using a proposed mean and median feature, which is extracted at least factor 2.5 faster than alternative features with comparable performance.

We developed a novel electronic cleansing (EC) method for dual-energy CT colonography (DE-CTC) based on an ensemble deep convolution neural network (DCNN) and multi-spectral multi-slice image patches. In the method, an ensemble DCNN is used to classify each voxel of a DE-CTC image volume into five classes: luminal air, soft tissue, tagged fecal materials, and partial-volume boundaries between air and tagging and those between soft tissue and tagging. Each DCNN acts as a voxel classifier, where an input image patch centered at the voxel is generated as input to the DCNNs. An image patch has three channels that are mapped from a region-of-interest containing the image plane of the voxel and the two adjacent image planes. Six different types of spectral input image datasets were derived using two dual-energy CT images, two virtual monochromatic images, and two material images. An ensemble DCNN was constructed by use of a meta-classifier that combines the output of multiple DCNNs, each of which was trained with a different type of multi-spectral image patches. The electronically cleansed CTC images were calculated by removal of regions classified as other than soft tissue, followed by a colon surface reconstruction. For pilot evaluation, 359 volumes of interest (VOIs) representing sources of subtraction artifacts observed in current EC schemes were sampled from 30 clinical CTC cases. Preliminary results showed that the ensemble DCNN can yield high accuracy in labeling of the VOIs, indicating that deep learning of multi-spectral EC with multi-slice imaging could accurately remove residual fecal materials from CTC images without generating major EC artifacts.

Colorectal cancer (CRC) is one of the most common and deadliest forms of cancer, accounting for nearly 10% of all forms of cancer in the world. Even though colonoscopy is considered the most effective method for screening and diagnosis, the success of the procedure is highly dependent on the operator skills and level of hand-eye coordination. In this work, we propose to adapt fully convolution neural networks (FCN), to identify and segment polyps in colonoscopy images. We converted three established networks into a fully convolution architecture and fine-tuned their learned representations to the polyp segmentation task. We validate our framework on the 2015 MICCAI polyp detection challenge dataset, surpassing the state-of-the-art in automated polyp detection. Our method obtained high segmentation accuracy and a detection precision and recall of 73.61% and 86.31%, respectively.

Robust training of a deep convolutional neural network (DCNN) requires a very large number of annotated datasets that are currently not available in CT colonography (CTC). We previously demonstrated that deep transfer learning provides an effective approach for robust application of a DCNN in CTC. However, at high detection accuracy, the differentiation of small polyps from non-polyps was still challenging. In this study, we developed and evaluated a deep ensemble learning (DEL) scheme for reviewing of virtual endoluminal images to improve the performance of computer-aided detection (CADe) of polyps in CTC. Nine different types of image renderings were generated from virtual endoluminal images of polyp candidates detected by a conventional CADe system. Eleven DCNNs that represented three types of publically available pre-trained DCNN models were re-trained by transfer learning to identify polyps from the virtual endoluminal images. A DEL scheme that determines the final detected polyps by a review of the nine types of VE images was developed by combining the DCNNs using a random forest classifier as a meta-classifier. For evaluation, we sampled 154 CTC cases from a large CTC screening trial and divided the cases randomly into a training dataset and a test dataset. At 3.9 falsepositive (FP) detections per patient on average, the detection sensitivities of the conventional CADe system, the highestperforming single DCNN, and the DEL scheme were 81.3%, 90.7%, and 93.5%, respectively, for polyps ≥6 mm in size. For small polyps, the DEL scheme reduced the number of false positives by up to 83% over that of using a single DCNN alone. These preliminary results indicate that the DEL scheme provides an effective approach for improving the polyp detection performance of CADe in CTC, especially for small polyps.

Serrated polyps were previously believed to be benign lesions with no cancer potential. However, recent studies have revealed a novel molecular pathway where also serrated polyps can develop into colorectal cancer. CT colonography (CTC) can detect serrated polyps using the radiomic biomarker of contrast coating, but this requires expertise from the reader and current computer-aided detection (CADe) systems have not been designed to detect the contrast coating. The purpose of this study was to develop a novel CADe method that makes use of deep learning to detect serrated polyps based on their contrast-coating biomarker in CTC. In the method, volumetric shape-based features are used to detect polyp sites over soft-tissue and fecal-tagging surfaces of the colon. The detected sites are imaged using multi-angular 2D image patches. A deep convolutional neural network (DCNN) is used to review the image patches for the presence of polyps. The DCNN-based polyp-likelihood estimates are merged into an aggregate likelihood index where highest values indicate the presence of a polyp. For pilot evaluation, the proposed DCNN-CADe method was evaluated with a 10-fold cross-validation scheme using 101 colonoscopy-confirmed cases with 144 biopsy-confirmed serrated polyps from a CTC screening program, where the patients had been prepared for CTC with saline laxative and fecal tagging by barium and iodine-based diatrizoate. The average per-polyp sensitivity for serrated polyps ≥6 mm in size was 93±7% at 0:8±1:8 false positives per patient on average. The detection accuracy was substantially higher that of a conventional CADe system. Our results indicate that serrated polyps can be detected automatically at high accuracy in CTC.

Accurate identification of polyps is the ultimate goal of Computed Tomography Colonography (CTC). While oral contrast agents were originally used to tag stool and fluid for the ultimate goal of CTC, recently their effect on coating the surface of polyps has been observed. This study aims to evaluate (1) the frequency at which the oral contrast adhered to polyp surfaces and (2) if there was a difference in contrast adherence with respect to diverse polyp types. To eliminate gravity as a factor in this study, the polyps in contact with tagged fluid pools, particularly on the bottom of the colon wall were excluded. A total of 150 polyps were selected under the above condition from a CTC database and screened for any adherent contrast on the luminal edge. Among the total, 53% of the screened polyps had adherent contrast. Serrated adenomas and hyperplastic polyps had a higher tagging percentage, 77.80% and 62.50% respectively, than tubular adenomas and tubulovillous adenomas, 44.40% and 43% respectively. Other factors that were analyzed for the effect on coating include size and location of the polyps. The higher tagging percentage of serrated adenomas and hyperplastic polyps may be due to their similar cellular features. The average size of the polyps was 8.9 mm. When the polyps were separated by size into small (5-9mm) and large (10-26mm) groups, the large group had a higher tagging percentage. The polyp types were also classified by location with the major findings being: 1) Tubular adenomas were present in all segments of the colon and 2) that serrated adenomas were present at a higher percentage in the proximal colon. These findings shall facilitate characterizing tagging agents and improve computer aided detection and classification of polyps via CTC.

We propose a method for computer-assisted optical biopsy for colorectal polyps, with the final goal of assisting the medical expert during the colonoscopy. In particular, we target the problem of automatic classification of polyp images in two classes: adenomatous vs non-adenoma. Our approach is based on recent advancements in convolutional neural networks (CNN) for image representation. In the paper, we describe and compare four different methodologies to address the binary classification task: a baseline with classical features and a Random Forest classifier, two methods based on features obtained from a pre-trained network, and finally, the end-to-end training of a CNN. With the pre-trained network, we show the feasibility of transferring a feature extraction mechanism trained on millions of natural images, to the task of classifying adenomatous polyps. We then demonstrate further performance improvements when training the CNN for our specific classification task. In our study, 776 polyp images were acquired and histologically analyzed after polyp resection. We report a performance increase of the CNN-based approaches with respect to both, the conventional engineered features and to a state-of-the-art method based on videos and 3D shape features.

Cardiothoracic ratio (CTR) is a widely used radiographic index to assess heart size on chest X-rays (CXRs). Recent studies have suggested that also two-dimensional CTR might contain clinical information about the heart function. However, manual measurement of such indices is both subjective and time consuming. This study proposes a fast algorithm to automatically estimate CTR indices based on CXRs. The algorithm has three main steps: 1) model based lung segmentation, 2) estimation of heart boundaries from lung contours, and 3) computation of cardiothoracic indices from the estimated boundaries. We extended a previously employed lung detection algorithm to automatically estimate heart boundaries without using ground truth heart markings. We used two datasets: a publicly available dataset with 247 images as well as clinical dataset with 167 studies from Geisinger Health System. The models of lung fields are learned from both datasets. The lung regions in a given test image are estimated by registering the learned models to patient CXRs. Then, heart region is estimated by applying Harris operator on segmented lung fields to detect the corner points corresponding to the heart boundaries. The algorithm calculates three indices, CTR1D, CTR2D, and cardiothoracic area ratio (CTAR). The method was tested on 103 clinical CXRs and average error rates of 7.9%, 25.5%, and 26.4% (for CTR1D, CTR2D, and CTAR respectively) were achieved. The proposed method outperforms previous CTR estimation methods without using any heart templates. This method can have important clinical implications as it can provide fast and accurate estimate of cardiothoracic indices.

A fully automated computer algorithm has been developed to evaluate coronary artery calcification (CAC) from lowdose CT scans. CAC is identified and evaluated in three main coronary artery groups: Left Main and Left Anterior Descending Artery (LM + LAD) CAC, Left Circumflex Artery (LCX) CAC, and Right Coronary Artery (RCA) CAC. The artery labeling is achieved by segmenting all CAC candidates in the heart region and applying geometric constraints on the candidates using locally pre-identified anatomy regions. This algorithm was evaluated on 1,359 low-dose ungated CT scans, in which each artery CAC content was categorically visually scored by a radiologist into none, mild, moderate and extensive. The Spearman correlation coefficient R was used to assess the agreement between three automated CAC scores (Agatston-weighted, volume, and mass) and categorical visual scores. For Agatston-weighted automated scores, R was 0.87 for total CAC, 0.82 for LM + LAD CAC, 0.66 for LCX CAC and 0.72 for RCA CAC; results using volume and mass scores were similar. CAC detection sensitivities were: 0.87 for total, 0.82 for LM + LAD, 0.65 for LCX and 0.74 for RCA. To assess the impact of image noise, the dataset was further partitioned into three subsets based on heart region noise level (low<=80HU, medium=(80HU, 110HU], high>110HU). The low and medium noise subsets had higher sensitivities and correlations than the high noise subset. These results indicate that location specific heart risk assessment is possible from low-dose chest CT images.

Coronary artery calcification (CAC) is a typical marker of the coronary artery disease, which is one of the biggest causes of mortality in the U.S. This study evaluates the feasibility of using a deep convolutional neural network (DCNN) to automatically detect CAC in X-ray images. 1768 posteroanterior (PA) view chest X-Ray images from Sichuan Province Peoples Hospital, China were collected retrospectively. Each image is associated with a corresponding diagnostic report written by a trained radiologist (907 normal, 861 diagnosed with CAC). Onequarter of the images were randomly selected as test samples; the rest were used as training samples. DCNN models consisting of 2,4,6 and 8 convolutional layers were designed using blocks of pre-designed CNN layers. Each block was implemented in Theano with Graphics Processing Units (GPU). Human-in-the-loop learning was also performed on a subset of 165 images with framed arteries by trained physicians. The results from the DCNN models were compared to the diagnostic reports. The average diagnostic accuracies for models with 2,4,6,8 layers were 0.85, 0.87, 0.88, and 0.89 respectively. The areas under the curve (AUC) were 0.92, 0.95, 0.95, and 0.96. As the model grows deeper, the AUC or diagnostic accuracies did not have statistically significant changes. The results of this study indicate that DCNN models have promising potential in the field of intelligent medical image diagnosis practice.

Flow Doppler imaging is widely used by clinicians to detect diseases of the valves. In particular, continuous wave (CW) Doppler mode scan is routinely done during echocardiography and shows Doppler signal traces over multiple heart cycles. Traditionally, echocardiographers have manually traced such velocity envelopes to extract measurements such as decay time and pressure gradient which are then matched to normal and abnormal values based on clinical guidelines. In this paper, we present a fully automatic approach to deriving these measurements for aortic stenosis retrospectively from echocardiography videos. Comparison of our method with measurements made by echocardiographers shows large agreement as well as identification of new cases missed by echocardiographers.

Late Gadolinium-Enhanced Cardiac MRI (LGE CMRI) is an emerging non-invasive technique to image and quantify preablation native and post-ablation atrial scarring. Previous studies have reported that enhanced image intensities of the atrial scarring in the LGE CMRI inversely correlate with the left atrial endocardial voltage invasively obtained by electro-anatomical mapping. However, the reported reproducibility of using LGE CMRI to identify and quantify atrial scarring is variable. This may be due to two reasons: first, delineation of the left atrium (LA) and pulmonary veins (PVs) anatomy generally relies on manual operation that is highly subjective, and this could substantially affect the subsequent atrial scarring segmentation; second, simple intensity based image features may not be good enough to detect subtle changes in atrial scarring. In this study, we hypothesized that texture analysis can provide reliable image features for the LGE CMRI images subject to accurate and objective delineation of the heart anatomy based on a fully-automated whole heart segmentation (WHS) method. We tested the extracted texture features to differentiate between pre-ablation and post-ablation LGE CMRI studies in longstanding persistent atrial fibrillation patients. These patients often have extensive native scarring and differentiation from post-ablation scarring can be difficult. Quantification results showed that our method is capable of solving this classification task, and we can envisage further deployment of this texture analysis based method for other clinical problems using LGE CMRI.

Screening mammography is the most effective and low-cost method to date for early cancer detection. Mammographic breast density has been shown to be highly correlated with breast cancer risk. We are developing a radiomic model for BI-RADS density categorization on digital mammography (FFDM) with a supervised machine learning approach. With IRB approval, we retrospectively collected 478 FFDMs from 478 women. As a gold standard, breast density was assessed by an MQSA radiologist based on BI-RADS categories. The raw FFDMs were used for computerized density assessment. The raw FFDM first underwent log-transform to approximate the x-ray sensitometric response, followed by multiscale processing to enhance the fibroglandular densities and parenchymal patterns. Three ROIs were automatically identified based on the keypoint distribution, where the keypoints were obtained as the extrema in the image Gaussian scale-space. A total of 73 features, including intensity and texture features that describe the density and the parenchymal pattern, were extracted from each breast. Our BI-RADS density estimator was constructed by using a random forest classifier. We used a 10-fold cross validation resampling approach to estimate the errors. With the random forest classifier, computerized density categories for 412 of the 478 cases agree with radiologist’s assessment (weighted kappa = 0.93). The machine learning method with radiomic features as predictors demonstrated a high accuracy in classifying FFDMs into BI-RADS density categories. Further work is underway to improve our system performance as well as to perform an independent testing using a large unseen FFDM set.

We proposed the neutrosophic approach for segmenting breast lesions in breast Computer Tomography (bCT) images. The neutrosophic set (NS) considers the nature and properties of neutrality (or indeterminacy), which is neither true nor false. We considered the image noise as an indeterminate component, while treating the breast lesion and other breast areas as true and false components. We first transformed the image into the NS domain. Each voxel in the image can be described as its membership in True, Indeterminate, and False sets. Operations α-mean, β-enhancement, and γ-plateau iteratively smooth and contrast-enhance the image to reduce the noise level of the true set. Once the true image no longer changes, we applied one existing algorithm for bCT images, the RGI segmentation, on the resulting image to segment the breast lesions. We compared the segmentation performance of the proposed method (named as NS-RGI) to that of the regular RGI segmentation. We used a total of 122 breast lesions (44 benign, 78 malignant) of 123 non-contrasted bCT cases. We measured the segmentation performances of the NS-RGI and the RGI using the DICE coefficient. The average DICE value of the NS-RGI was 0.82 (STD: 0.09), while that of the RGI was 0.8 (STD: 0.12). The difference between the two DICE values was statistically significant (paired t test, p-value = 0.0007). We conducted a subsequent feature analysis on the resulting segmentations. The classifier performance for the NS-RGI (AUC = 0.8) improved over that of the RGI (AUC = 0.69, p-value = 0.006).

Breast cancer is the most common cancer diagnosed among US women and the second leading cause of cancer death 1 . Breast density is an independent risk factor for breast cancer and more than 25 states mandate its reporting to patients as part of the lay mammogram report ² . Recent publications have demonstrated that breast density measured from low-dose chest CT (LDCT) correlates well with that measured from mammograms and MRIs ^3-4 , thereby providing valuable information for many women who have undergone LDCT but not recent mammograms. A fully automated framework for breast density assessment from LDCT is presented in this paper. The whole breast region is first segmented using an anatomy-orientated novel approach based on the propagation of muscle fronts for separating the fibroglandular tissue from the underlying muscles. The fibroglandular tissue regions are then identified from the segmented whole breast and the percentage density is calculated based on the volume ratio of the fibroglandular tissue to the local whole breast region. The breast region segmentation framework was validated with 1270 LDCT scans, with 96.1% satisfactory outcomes based on visual inspection. The density assessment was evaluated by comparing with BI-RADS density grades established by an experienced radiologist in 100 randomly selected LDCT scans of female subjects. The continuous breast density measurement was shown to be consistent with the reference subjective grading, with the Spearman’s rank correlation 0.91 (p-value < 0.001). After converting the continuous density to categorical grades, the automated density assessment was congruous with the radiologist’s reading in 91% cases.

We assess the feasibility of a parenchymal texture feature fusion approach, utilizing a convolutional neural network (ConvNet) architecture, to benefit breast cancer risk assessment. Hypothesizing that by capturing sparse, subtle interactions between localized motifs present in two-dimensional texture feature maps derived from mammographic images, a multitude of texture feature descriptors can be optimally reduced to five meta-features capable of serving as a basis on which a linear classifier, such as logistic regression, can efficiently assess breast cancer risk. We combine this methodology with our previously validated lattice-based strategy for parenchymal texture analysis and we evaluate the feasibility of this approach in a case-control study with 424 digital mammograms. In a randomized split-sample setting, we optimize our framework in training/validation sets (N=300) and evaluate its descriminatory performance in an independent test set (N=124). The discriminatory capacity is assessed in terms of the the area under the curve (AUC) of the receiver operator characteristic (ROC). The resulting meta-features exhibited strong classification capability in the test dataset (AUC = 0.90), outperforming conventional, non-fused, texture analysis which previously resulted in an AUC=0.85 on the same case-control dataset. Our results suggest that informative interactions between localized motifs exist and can be extracted and summarized via a fairly simple ConvNet architecture.

Understanding the key radiogenomic associations for breast cancer between DCE-MRI and micro-RNA expressions is the foundation for the discovery of radiomic features as biomarkers for assessing tumor progression and prognosis. We conducted a study to analyze the radiogenomic associations for breast cancer using the TCGA-TCIA data set. The core idea that tumor etiology is a function of the behavior of miRNAs is used to build the regression models. The associations based on regression are analyzed for three study outcomes: diagnosis, prognosis, and treatment. The diagnosis group consists of miRNAs associated with clinicopathologic features of breast cancer and significant aberration of expression in breast cancer patients. The prognosis group consists of miRNAs which are closely associated with tumor suppression and regulation of cell proliferation and differentiation. The treatment group consists of miRNAs that contribute significantly to the regulation of metastasis thereby having the potential to be part of therapeutic mechanisms. As a first step, important miRNA expressions were identified and their ability to classify the clinical phenotypes based on the study outcomes was evaluated using the area under the ROC curve (AUC) as a figure-of-merit. The key mapping between the selected miRNAs and radiomic features were determined using least absolute shrinkage and selection operator (LASSO) regression analysis within a two-loop leave-one-out cross-validation strategy. These key associations indicated a number of radiomic features from DCE-MRI to be potential biomarkers for the three study outcomes.

DCE-MRI datasets have a temporal aspect to them, resulting in multiple regions of interest (ROIs) per subject, based on contrast time points. It is unclear how the different contrast time points vary in terms of usefulness for computer-aided diagnosis tasks in conjunction with deep learning methods. We thus sought to compare the different DCE-MRI contrast time points with regard to how well their extracted features predict response to neoadjuvant chemotherapy within a deep convolutional neural network. Our dataset consisted of 561 ROIs from 64 subjects. Each subject was categorized as a non-responder or responder, determined by recurrence-free survival. First, features were extracted from each ROI using a convolutional neural network (CNN) pre-trained on non-medical images. Linear discriminant analysis classifiers were then trained on varying subsets of these features, based on their contrast time points of origin. Leave-one-out cross validation (by subject) was used to assess performance in the task of estimating probability of response to therapy, with area under the ROC curve (AUC) as the metric. The classifier trained on features from strictly the pre-contrast time point performed the best, with an AUC of 0.85 (SD = 0.033). The remaining classifiers resulted in AUCs ranging from 0.71 (SD = 0.028) to 0.82 (SD = 0.027). Overall, we found the pre-contrast time point to be the most effective at predicting response to therapy and that including additional contrast time points moderately reduces variance.

Chromoendoscopy (CH) is a gastroenterology imaging modality that involves the staining of tissues with methylene blue, which reacts with the internal walls of the gastrointestinal tract, improving the visual contrast in mucosal surfaces and thus enhancing a doctor’s ability to screen precancerous lesions or early cancer. This technique helps identify areas that can be targeted for biopsy or treatment and in this work we will focus on gastric cancer detection. Gastric chromoendoscopy for cancer detection has several taxonomies available, one of which classifies CH images into three classes (normal, metaplasia, dysplasia) based on color, shape and regularity of pit patterns. Computer-assisted diagnosis is desirable to help us improve the reliability of the tissue classification and abnormalities detection. However, traditional computer vision methodologies, mainly segmentation, do not translate well to the specific visual characteristics of a gastroenterology imaging scenario. We propose the exploitation of a first unsupervised segmentation via superpixel, which groups pixels into perceptually meaningful atomic regions, used to replace the rigid structure of the pixel grid. For each superpixel, a set of features is extracted and then fed to a random forest based classifier, which computes a model used to predict the class of each superpixel. The average general accuracy of our model is 92.05% in the pixel domain (86.62% in the superpixel domain), while detection accuracies on the normal and abnormal class are respectively 85.71% and 95%. Eventually, the whole image class can be predicted image through a majority vote on each superpixel's predicted class.

Pancreatic cancer is the most lethal cancer with an overall 5-year survival rate of 7%¹ due to the late stage at diagnosis and the ineffectiveness of current therapeutic strategies. Given the poor prognosis, early detection at a pre-cancerous stage is the best tool for preventing this disease. Intraductal papillary mucinous neoplasms (IPMN), cystic tumors of the pancreas, represent the only radiographically identifiable precursor lesion of pancreatic cancer and are known to evolve stepwise from low-to-high-grade dysplasia before progressing into an invasive carcinoma. Observation is usually recommended for low-risk (low- and intermediate-grade dysplasia) patients, while high-risk (high-grade dysplasia and invasive carcinoma) patients undergo resection; hence, patient selection is critically important in the management of pancreatic cysts.² Radiologists use standard criteria such as main pancreatic duct size, cyst size, or presence of a solid enhancing component in the cyst to optimally select patients for surgery.³ However, these findings are subject to a radiologist’s interpretation and have been shown to be inconsistent with regards to the presence of a mural nodule or solid component.⁴ We propose objective classification of risk groups based on quantitative imaging features extracted from CT scans. We apply new features that represent the solid component (i.e. areas of high intensity) within the cyst and extract standard texture features. An adaptive boost classifier⁵ achieves the best performance with area under receiver operating characteristic curve (AUC) of 0.73 and accuracy of 77.3% for texture features. The random forest classifier achieves the best performance with AUC of 0.71 and accuracy of 70.8% with the solid component features.

Magnetic Resonance Elastography (MRE) is a phase-contrast MRI technique which calculates quantitative stiffness images, called elastograms, by imaging the propagation of acoustic waves in tissues. It is used clinically to diagnose liver fibrosis. Automated analysis of MRE is difficult as the corresponding MRI magnitude images (which contain anatomical information) are affected by intensity inhomogeneity, motion artifact, and poor tissue- and edge-contrast. Additionally, areas with low wave amplitude must be excluded. An automated algorithm has already been successfully developed and validated for clinical 2D MRE. 3D MRE acquires substantially more data and, due to accelerated acquisition, has exacerbated image artifacts. Also, the current 3D MRE processing does not yield a confidence map to indicate MRE wave quality and guide ROI selection, as is the case in 2D. In this study, extension of the 2D automated method, with a simple wave-amplitude metric, was developed and validated against an expert reader in a set of 57 patient exams with both 2D and 3D MRE. The stiffness discrepancy with the expert for 3D MRE was -0.8% ± 9.45% and was better than discrepancy with the same reader for 2D MRE (-3.2% ± 10.43%), and better than the inter-reader discrepancy observed in previous studies. There were no automated processing failures in this dataset. Thus, the automated liver elasticity calculation (ALEC) algorithm is able to calculate stiffness from 3D MRE data with minimal bias and good precision, while enabling stiffness measurements to be fully reproducible and to be easily performed on the large 3D MRE datasets.

Tumor volume measured from computed tomography images is considered a biomarker for disease progression or treatment response. The estimation of the tumor volume depends on the imaging system parameters selected, as well as lesion characteristics. In this study, we examined how different image reconstruction methods affect the measurement of lesions in an anthropomorphic liver phantom with a non-uniform background. Iterative statistics-based and model-based reconstructions, as well as filtered back-projection, were evaluated and compared in this study. Statistics-based and filtered back-projection yielded similar estimation performance, while model-based yielded higher precision but lower accuracy in the case of small lesions. Iterative reconstructions exhibited higher signal-to-noise ratio but slightly lower contrast of the lesion relative to the background. A better understanding of lesion volumetry performance as a function of acquisition parameters and lesion characteristics can lead to its incorporation as a routine sizing tool.

Hepatocellular carcinoma (HCC) is the most common liver cancer and the third leading cause of cancer-related death worldwide.¹ Resection or liver transplantation may be curative in patients with early-stage HCC but early recurrence is common.^{2, 3} Microvascular invasion (MVI) is one of the most important predictors of early recurrence.³ The identification of MVI prior to surgery would optimally select patients for potentially curative resection or liver transplant. However, MVI can only be diagnosed by microscopic assessment of the resected tumor. The aim of the present study is to apply CT-based texture analysis to identify pre-operative imaging predictors of MVI in patients with HCC. Texture features are derived from CT and analyzed individually as well as in combination, to evaluate their ability to predict MVI. A two-stage classification is employed: HCC tumors are automatically categorized into uniform or heterogenous groups followed by classification into the presence or absence of MVI. We achieve an area under the receiver operating characteristic curve (AUC) of 0.76 and accuracy of 76.7% for uniform lesions and AUC of 0.79 and accuracy of 74.06% for heterogeneous tumors. These results suggest that MVI can be accurately and objectively predicted from preoperative CT scans.

Knee Osteoarthritis (OA) is the most common form of arthritis, affecting millions of people around the world. The effects of the disease have been studied using the shape and texture features of bones in PosteriorAnterior (PA) and Lateral radiographs separately. In this work we compare the utility of features from each view, and evaluate whether combining features from both is advantageous. We built a fully automated system to independently locate landmark points in both radiographic images using Random Forest Constrained Local Models. We extracted discriminative features from the two bony outlines using Appearance Models. The features were used to train Random Forest classifiers to solve three specific tasks: (i) OA classification, distinguishing patients with structural signs of OA from the others; (ii) predicting future onset of the disease and (iii) predicting which patients with no current pain will have a positive pain score later in a follow-up visit. Using a subset of the MOST dataset we show that the PA view has more discriminative features to classify and predict OA, while the lateral view contains features that achieve better performance in predicting pain, and that combining the features from both views gives a small improvement in accuracy of the classification compared to the individual views.

The effectiveness of phase contrast X-ray computed tomography (PCI-CT) in visualizing human patellar cartilage matrix has been demonstrated due to its ability to capture soft tissue contrast on a micrometer resolution scale. Recent studies have shown that off-the-shelf Convolutional Neural Network (CNN) features learned from a nonmedical data set can be used for medical image classification. In this paper, we investigate the ability of features extracted from two different CNNs for characterizing chondrocyte patterns in the cartilage matrix. We obtained features from 842 regions of interest annotated on PCI-CT images of human patellar cartilage using CaffeNet and Inception-v3 Network, which were then used in a machine learning task involving support vector machines with radial basis function kernel to classify the ROIs as healthy or osteoarthritic. Classification performance was evaluated using the area (AUC) under the Receiver Operating Characteristic (ROC) curve. The best classification performance was observed with features from Inception-v3 network (AUC = 0.95), which outperforms features extracted from CaffeNet (AUC = 0.91). These results suggest that such characterization of chondrocyte patterns using features from internal layers of CNNs can be used to distinguish between healthy and osteoarthritic tissue with high accuracy.

Few studies have analyzed the microstructural properties of bone in cases of Osteogenenis Imperfecta (OI), or ‘brittle bone disease’. Current approaches mainly focus on bone mineral density measurements as an indirect indicator of bone strength and quality. It has been shown that bone strength would depend not only on composition but also structural organization. This study aims to characterize 3D structure of the cortical bone in high-resolution micro CT images. A total of 40 bone fragments from 28 subjects (13 with OI and 15 healthy controls) were imaged using micro tomography using a synchrotron light source (SRµCT). Minkowski functionals - volume, surface, curvature, and Euler characteristics - describing the topological organization of the bone were computed from the images. The features were used in a machine learning task to classify between healthy and OI bone. The best classification performance (mean AUC – 0.96) was achieved with a combined 4-dimensional feature of all Minkowski functionals. Individually, the best feature performance was seen using curvature (mean AUC - 0.85), which characterizes the edges within a binary object. These results show that quantitative analysis of cortical bone microstructure, in a computer-aided diagnostics framework, can be used to distinguish between healthy and OI bone with high accuracy.

Melanoma is considered a major health problem since it is the deadliest form of skin cancer. The early diagnosis through periodic screening with dermoscopic images can significantly improve the survival rate as well as reduce the treatment cost and consequent suffering of patients. Dermoscopy or skin surface microscopy provides in vivo inspection of color and morphologic structures of pigmented skin lesions (PSLs), rendering higher accuracy for detecting suspicious cases than it is possible via inspecting with naked eye. However, interpretation of dermoscopic images is time consuming and subjective, even for trained dermatologists. Therefore, there is currently a great interest in the development of computeraided diagnosis (CAD) systems for automated melanoma recognition. However, the majority of the CAD systems are still in the early development stage with lack of descriptive feature generation and benchmark evaluation in ground-truth datasets. This work is focusing on by addressing the various issues related to the development of such a CAD system with effective feature extraction from Non-Subsampled Contourlet Transform (NSCT) and Eig(Hess) histogram of oriented gradients (HOG) and lesion classification with efficient Extreme Learning Machine (ELM) due to its good generalization abilities and a high learning efficiency and evaluating its effectiveness in a benchmark data set of dermoscopic images towards the goal of realistic comparison and real clinical integration. The proposed research on melanoma recognition has huge potential for offering powerful services that would significantly benefit the present Biomedical Information Systems.

The abstract is not available

A convolutional network architecture is presented to determine bounding boxes around six organs in thoraxabdomen CT scans. A single network for each orthogonal view determines the presence of lungs, kidneys, spleen and liver. We show that an architecture that takes additional slices before and after the slice of interest as an additional input outperforms an architecture that processes single slices. From the slice-based analysis, a bounding box around the structures of interest can be computed. The system uses 6 convolutional, 4 pooling and one fully connected layer and uses 333 scans for training and 110 for validation. The test set contains 110 scans. The average Dice score of the proposed method was 0.95 and 0.95 for the lungs, 0.59 and 0.58 for the kidneys, 0.83 for the liver and 0.63 for the spleen. This paper shows that automatic localization of organs using multi-label convolution neural networks is possible. This architecture can likely be used to identify other organs of interest as well.

Image-derived features of breast parenchymal texture patterns have emerged as promising risk factors for breast cancer, paving the way towards personalized recommendations regarding women’s cancer risk evaluation and screening. The main steps to extract texture features of the breast parenchyma are the selection of regions of interest (ROIs) where texture analysis is performed, the texture feature calculation and the texture feature summarization in case of multiple ROIs. In this study, we incorporate breast anatomy in these three key steps by (a) introducing breast anatomical sampling for the definition of ROIs, (b) texture feature calculation aligned with the structure of the breast and (c) weighted texture feature summarization considering the spatial position and the underlying tissue composition of each ROI. We systematically optimize this novel framework for parenchymal tissue characterization in a case-control study with digital mammograms from 424 women. We also compare the proposed approach with a conventional methodology, not considering breast anatomy, recently shown to enhance the case-control discriminatory capacity of parenchymal texture analysis. The case-control classification performance is assessed using elastic-net regression with 5-fold cross validation, where the evaluation measure is the area under the curve (AUC) of the receiver operating characteristic. Upon optimization, the proposed breast-anatomy-driven approach demonstrated a promising case-control classification performance (AUC=0.87). In the same dataset, the performance of conventional texture characterization was found to be significantly lower (AUC=0.80, DeLong's test p-value<0.05). Our results suggest that breast anatomy may further leverage the associations of parenchymal texture features with breast cancer, and may therefore be a valuable addition in pipelines aiming to elucidate quantitative mammographic phenotypes of breast cancer risk.

The application of endovascular aortic aneurysm repair has expanded over the last decade. However, the long-term performance of stent grafts, in particular durable fixation and sealing to the aortic wall, remains the main concern of this treatment. The sealing and fixation are challenged at every heartbeat due to downward and radial pulsatile forces. Yet knowledge on cardiac-induced dynamics of implanted stent grafts is sparse, as it is not measured in routine clinical follow-up. Such knowledge is particularly relevant to perform fatigue tests, to predict failure in the individual patient and to improve stent graft designs. Using a physical dynamic stent graft model in an anthropomorphic phantom, we have evaluated the performance of our previously proposed segmentation and registration algorithm to detect periodic motion of stent grafts on ECG-gated (3D+t) CT data. Abdominal aortic motion profiles were simulated in two series of Gaussian based patterns with different amplitudes and frequencies. Experiments were performed on a 64-slice CT scanner with a helical scan protocol and retrospective gating. Motion patterns as estimated by our algorithm were compared to motion patterns obtained from optical camera recordings of the physical stent graft model in motion. Absolute errors of the patterns' amplitude were smaller than 0.28 mm. Even the motion pattern with an amplitude of 0.23 mm was measured, although the amplitude of motion was overestimated by the algorithm with 43%. We conclude that the algorithm performs well for measurement of stent graft motion in the mm and sub-mm range. This ultimately is expected to aid in patient-specific risk assessment and improving stent graft designs.

Hereditary hemorrhagic telangiectasia (HHT) is an autosomic dominant disorder, which is characterized by the development of multiple arterio-venous malformations in the skin, mucous membranes, and/or visceral organs. Pulmonary Arterio-Venous Malformation (PAVM) is an abnormal connection where feeding arteries shunt directly into draining veins with no intervening capillary bed. This condition may lead to paradoxical embolism and hemorrhagic complications. PAVMs patients should systematically be screened as the spontaneous complication rate is high, reaching almost 50%. Chest enhanced contrast CT scanner is the reference screening and follow-up examination. When performed by experienced operators as the prime treatment, percutaneous embolization of PAVMs is a safe, efficient and sustained therapy. The accuracy of PAVM detection and quantification of its progression over time is the key of embolotherapy success. In this paper, we propose an automatic method for PAVM detection and quantification relying on a modeling of vessel deformation, i.e. local caliber increase, based on mathematical morphology. The pulmonary field and vessels are first segmented using geodesic operators. The vessel caliber is estimated by means of a granulometric measure and the local caliber increase is detected by using a geodesic operator, the h-maxdomes. The detection sensitivity can be tuned up according to the choice of the h value which models the irregularity of the vessel caliber along its axis and the PAVM selection is performed according to a clinical criterion of >3 mm diameter of the feeding artery of the PAVM. The developed method was tested on a 20 patient dataset. A sensitivity study allowed choosing the irregularity parameter to maximize the true positive ratio reaching 85.4% in average. A specific false positive reduction procedure targeting the vessel trunks of the arterio-venous tree near mediastinum allows a precision increase from 13% to 67% with an average number of 1.15 false positives per scan.

Artery calcification is observed commonly in elderly patients, especially in patients with chronic kidney disease, and may affect coronary, carotid and peripheral arteries. Vascular calcification has been associated with many clinical outcomes. Manual identification of calcification in CT scans requires substantial expert interaction, which makes it time-consuming and infeasible for large-scale studies. Many works have been proposed for coronary artery calcification detection in cardiac CT scans. In these works, coronary artery extraction is commonly required for calcification detection. However, there are few works about abdominal or pelvic artery calcification detection. In this work, we present a method for automatic pelvic artery calcification detection on CT scan. This method uses the recent advanced faster region-based convolutional neural network (R-CNN) to directly identify artery calcification without a need for artery extraction since pelvic artery extraction itself is challenging. Our method first generates category-independent region proposals for each slice of the input CT scan using region proposal networks (RPN). Then, each region proposal is jointly classified and refined by softmax classifier and bounding box regressor. We applied the detection method to 500 images from 20 CT scans of patients for evaluation. The detection system achieved a 77.4% average precision and a 85% sensitivity at 1 false positive per image.

4D arterial spin labeling magnetic resonance angiography (4D ASL MRA) is a non-invasive and safe modality for cerebrovascular imaging procedures. It uses the patient’s magnetically labeled blood as intrinsic contrast agent, so that no external contrast media is required. It provides important 3D structure and blood flow information but a sufficient cerebrovascular segmentation is important since it can help clinicians to analyze and diagnose vascular diseases faster, and with higher confidence as compared to simple visual rating of raw ASL MRA images. This work presents a new method for automatic cerebrovascular segmentation in 4D ASL MRA images of the brain. In this process images are denoised, corresponding image label/control image pairs of the 4D ASL MRA sequences are subtracted, and temporal intensity averaging is used to generate a static representation of the vascular system. After that, sets of vessel and background seeds are extracted and provided as input for the image foresting transform algorithm to segment the vascular system. Four 4D ASL MRA datasets of the brain arteries of healthy subjects and corresponding time-of-flight (TOF) MRA images were available for this preliminary study. For evaluation of the segmentation results of the proposed method, the cerebrovascular system was automatically segmented in the high-resolution TOF MRA images using a validated algorithm and the segmentation results were registered to the 4D ASL datasets. Corresponding segmentation pairs were compared using the Dice similarity coefficient (DSC). On average, a DSC of 0.9025 was achieved, indicating that vessels can be extracted successfully from 4D ASL MRA datasets by the proposed segmentation method.

Neurosurgeons currently base most of their treatment decisions for intracranial aneurysms (IAs) on morphological measurements made manually from 2D angiographic images. These measurements tend to be inaccurate because 2D measurements cannot capture the complex geometry of IAs and because manual measurements are variable depending on the clinician’s experience and opinion. Incorrect morphological measurements may lead to inappropriate treatment strategies. In order to improve the accuracy and consistency of morphological analysis of IAs, we have developed an image-based computational tool, AView. In this study, we quantified the accuracy of computer-assisted adjuncts of AView for aneurysmal morphologic assessment by performing measurement on spheres of known size and anatomical IA models. AView has an average morphological error of 0.56% in size and 2.1% in volume measurement. We also investigate the clinical utility of this tool on a retrospective clinical dataset and compare size and neck diameter measurement between 2D manual and 3D computer-assisted measurement. The average error was 22% and 30% in the manual measurement of size and aneurysm neck diameter, respectively. Inaccuracies due to manual measurements could therefore lead to wrong treatment decisions in 44% and inappropriate treatment strategies in 33% of the IAs. Furthermore, computer-assisted analysis of IAs improves the consistency in measurement among clinicians by 62% in size and 82% in neck diameter measurement. We conclude that AView dramatically improves accuracy for morphological analysis. These results illustrate the necessity of a computer-assisted approach for the morphological analysis of IAs.

In this paper, we propose a novel supervoxel segmentation method designed for mediastinal lymph node by embedding Hessian-based feature extraction. Starting from a popular supervoxel segmentation method, SLIC, which computes supervoxels by minimising differences of intensity and distance, we overcome this method's limitation of merging neighboring regions with similar intensity by introducing Hessian-based feature analysis into the supervoxel formation. We call this structure-oriented voxel clustering, which allows more accurate division into distinct regions having blob-, line- or sheet-like structures. This way, different tissue types in chest CT volumes can be segmented individually, even if neighboring tissues have similar intensity or are of non- spherical extent. We demonstrate the performance of the Hessian-assisted supervoxel technique by applying it to mediastinal lymph node detection in 47 chest CT volumes, resulting in false positive reductions from lymph node candidate regions. 89 % of lymph nodes whose short axis is at least 10 mm could be detected with 5.9 false positives per case using our method, compared to our previous method having 83 % of detection rate with 6.4 false positives per case.

Recent breakthroughs in training deep neural network architectures, in particular deep Convolutional Neural Networks (CNNs), made a big impact on vision research and are increasingly responsible for advances in Computer Aided Diagnosis (CAD). Since many natural scenes and medical images vary in size and are too large to feed to the networks as a whole, two stage systems are typically employed, where in the first stage, small regions of interest in the image are located and presented to the network as training and test data. These systems allow us to harness accurate region based annotations, making the problem easier to learn. However, information is processed purely locally and context is not taken into account. In this paper, we present preliminary work on the employment of a Conditional Random Field (CRF) that is trained on top the CNN to model contextual interactions such as the presence of other suspicious regions, for mammography CAD. The model can easily be extended to incorporate other sources of information, such as symmetry, temporal change and various patient covariates and is general in the sense that it can have application in other CAD problems.

We investigated whether deep learning has potential to aid in the diagnosis of breast cancer when applied to mammograms and biologic tissue composition images derived from three-compartment (3CB) imaging. The dataset contained diagnostic mammograms and 3CB images (water, lipid, and protein content) of biopsy-sampled BIRADS 4 and 5 lesions in 195 patients. In 58 patients, the lesion manifested as a mass (13 malignant vs. 45 benign), in 87 as microcalcifications (19 vs. 68), and in 56 as (focal) asymmetry or architectural distortion (11 vs. 45). Six patients had both a mass and calcifications. For each mammogram and corresponding 3CB images, a 128x128 region of interest containing the lesion was selected by an expert radiologist and used directly as input to a deep learning method pretrained on a very large independent set of non-medical images. We used a nested leave-one-out-by-case (patient) model selection and classification protocol. The area under the ROC curve (AUC) for the task of distinguishing between benign and malignant lesions was used as performance metric. For the cases with mammographic masses, the AUC increased from 0.83 (mammograms alone) to 0.89 (mammograms+3CB, p=.162). For the microcalcification and asymmetry/architectural distortion cases the AUC increased from 0.84 to 0.91 (p=.116) and from 0.61 to 0.87 (p=.006), respectively. Our results indicate great potential for the application of deep learning methods in the diagnosis of breast cancer and additional knowledge of the biologic tissue composition appeared to improve performance, especially for lesions mammographically manifesting as asymmetries or architectural distortions.

Intuitive segmentation-based CADx/radiomic features, calculated from the lesion segmentations of dynamic contrast-enhanced magnetic resonance images (DCE-MRIs) have been utilized in the task of distinguishing between malignant and benign lesions. Additionally, transfer learning with pre-trained deep convolutional neural networks (CNNs) allows for an alternative method of radiomics extraction, where the features are derived directly from the image data. However, the comparison of computer-extracted segmentation-based and CNN features in MRI breast lesion characterization has not yet been conducted. In our study, we used a DCE-MRI database of 640 breast cases – 191 benign and 449 malignant. Thirty-eight segmentation-based features were extracted automatically using our quantitative radiomics workstation. Also, 2D ROIs were selected around each lesion on the DCE-MRIs and directly input into a pre-trained CNN AlexNet, yielding CNN features. Each method was investigated separately and in combination in terms of performance in the task of distinguishing between benign and malignant lesions. Area under the ROC curve (AUC) served as the figure of merit. Both methods yielded promising classification performance with round-robin cross-validated AUC values of 0.88 (se =0.01) and 0.76 (se=0.02) for segmentationbased and deep learning methods, respectively. Combination of the two methods enhanced the performance in malignancy assessment resulting in an AUC value of 0.91 (se=0.01), a statistically significant improvement over the performance of the CNN method alone.

Diagnostic interpretation of breast MRI studies requires meticulous work and a high level of expertise. Computerized algorithms can assist radiologists by automatically characterizing the detected lesions. Deep learning approaches have shown promising results in natural image classification, but their applicability to medical imaging is limited by the shortage of large annotated training sets. In this work, we address automatic classification of breast MRI lesions using two different deep learning approaches. We propose a novel image representation for dynamic contrast enhanced (DCE) breast MRI lesions, which combines the morphological and kinetics information in a single multi-channel image. We compare two classification approaches for discriminating between benign and malignant lesions: training a designated convolutional neural network and using a pre-trained deep network to extract features for a shallow classifier. The domain-specific trained network provided higher classification accuracy, compared to the pre-trained model, with an area under the ROC curve of 0.91 versus 0.81, and an accuracy of 0.83 versus 0.71. Similar accuracy was achieved in classifying benign lesions, malignant lesions, and normal tissue images. The trained network was able to improve accuracy by using the multi-channel image representation, and was more robust to reductions in the size of the training set. A small-size convolutional neural network can learn to accurately classify findings in medical images using only a few hundred images from a few dozen patients. With sufficient data augmentation, such a network can be trained to outperform a pre-trained out-of-domain classifier. Developing domain-specific deep-learning models for medical imaging can facilitate technological advancements in computer-aided diagnosis.

Predicting the risk of occult invasive disease in ductal carcinoma in situ (DCIS) is an important task to help address the overdiagnosis and overtreatment problems associated with breast cancer. In this work, we investigated the feasibility of using computer-extracted mammographic features to predict occult invasive disease in patients with biopsy proven DCIS. We proposed a computer-vision algorithm based approach to extract mammographic features from magnification views of full field digital mammography (FFDM) for patients with DCIS. After an expert breast radiologist provided a region of interest (ROI) mask for the DCIS lesion, the proposed approach is able to segment individual microcalcifications (MCs), detect the boundary of the MC cluster (MCC), and extract 113 mammographic features from MCs and MCC within the ROI. In this study, we extracted mammographic features from 99 patients with DCIS (74 pure DCIS; 25 DCIS plus invasive disease). The predictive power of the mammographic features was demonstrated through binary classifications between pure DCIS and DCIS with invasive disease using linear discriminant analysis (LDA). Before classification, the minimum redundancy Maximum Relevance (mRMR) feature selection method was first applied to choose subsets of useful features. The generalization performance was assessed using Leave-One-Out Cross-Validation and Receiver Operating Characteristic (ROC) curve analysis. Using the computer-extracted mammographic features, the proposed model was able to distinguish DCIS with invasive disease from pure DCIS, with an average classification performance of AUC = 0.61 ± 0.05. Overall, the proposed computer-extracted mammographic features are promising for predicting occult invasive disease in DCIS.

When humans identify objects in images, context is an important cue; a cheetah is more likely to be a domestic cat when a television set is recognised in the background. Similar principles apply to the analysis of medical images. The detection of diseases that manifest unilaterally in symmetrical organs or organ pairs can in part be facilitated by a search for symmetrical discrepancies in or between the organs in question. During a mammographic exam, images are recorded of each breast and absence of a certain structure around the same location in the contralateral image will render the area under scrutiny more suspicious and conversely, the presence of similar tissue less so. In this paper, we present a fusion scheme for a deep Convolutional Neural Network (CNN) architecture with the goal to optimally capture such asymmetries. The method is applied to the domain of mammography CAD, but can be relevant to other medical image analysis tasks where symmetry is important such as lung, prostate or brain images.

Retinal vessel caliber changes are associated with several major diseases, such as diabetes and hypertension. These caliber changes can be evaluated using eye fundus images. However, the clinical assessment is tiresome and prone to errors, motivating the development of automatic methods. An automatic method based on vessel crosssection intensity profile model fitting for the estimation of vessel caliber in retinal images is herein proposed. First, vessels are segmented from the image, vessel centerlines are detected and individual segments are extracted and smoothed. Intensity profiles are extracted perpendicularly to the vessel, and the profile lengths are determined. Then, model fitting is applied to the smoothed profiles. A novel parametric model (DoG-L7) is used, consisting on a Difference-of-Gaussians multiplied by a line which is able to describe profile asymmetry. Finally, the parameters of the best-fit model are used for determining the vessel width through regression using ensembles of bagged regression trees with random sampling of the predictors (random forests). The method is evaluated on the REVIEW public dataset. A precision close to the observers is achieved, outperforming other state-of-the-art methods. The method is robust and reliable for width estimation in images with pathologies and artifacts, with performance independent of the range of diameters.

The Arteriolar-to-Venular Ratio (AVR) is a popular dimensionless measure which allows the assessment of patients’ condition for the early diagnosis of different diseases, including hypertension and diabetic retinopathy. This paper presents two new approaches for AVR computation in retinal photographs which include a sequence of automated processing steps: vessel segmentation, caliber measurement, optic disc segmentation, artery/vein classification, region of interest delineation, and AVR calculation. Both approaches have been tested on the INSPIRE-AVR dataset, and compared with a ground-truth provided by two medical specialists. The obtained results demonstrate the reliability of the fully automatic approach which provides AVR ratios very similar to at least one of the observers. Furthermore, the semi-automatic approach, which includes the manual modification of the artery/vein classification if needed, allows to significantly reduce the error to a level below the human error.

Diabetic retinopathy (DR) is a major cause of irreversible vision loss. DR screening relies on retinal clinical signs (features). Opportunities for computer-aided DR feature detection have emerged with the development of Ultra-WideField (UWF) digital scanning laser technology. UWF imaging covers 82% greater retinal area (200°), against 45° in conventional cameras³ , allowing more clinically relevant retinopathy to be detected⁴ . UWF images also provide a high resolution of 3078 x 2702 pixels. Currently DR screening uses 7 overlapping conventional fundus images, and the UWF images provide similar results^1,4. However, in 40% of cases, more retinopathy was found outside the 7-field ETDRS) fields by UWF and in 10% of cases, retinopathy was reclassified as more severe⁴ . This is because UWF imaging allows examination of both the central retina and more peripheral regions, with the latter implicated in DR⁶ . We have developed an algorithm for automatic recognition of DR features, including bright (cotton wool spots and exudates) and dark lesions (microaneurysms and blot, dot and flame haemorrhages) in UWF images. The algorithm extracts features from grayscale (green “red-free” laser light) and colour-composite UWF images, including intensity, Histogram-of-Gradient and Local binary patterns. Pixel-based classification is performed with three different classifiers. The main contribution is the automatic detection of DR features in the peripheral retina. The method is evaluated by leave-one-out cross-validation on 25 UWF retinal images with 167 bright lesions, and 61 other images with 1089 dark lesions. The SVM classifier performs best with AUC of 94.4% / 95.31% for bright / dark lesions.

Automated detection and quantification of spatio-temporal retinal changes is an important step to objectively assess disease progression and treatment effects for dynamic retinal diseases such as diabetic retinopathy (DR). However, detecting retinal changes caused by early DR lesions such as microaneurysms and dot hemorrhages from longitudinal pairs of fundus images is challenging due to intra and inter-image illumination variation between fundus images. This paper explores a method for automated detection of retinal changes from illumination normalized fundus images using a deep convolutional neural network (CNN), and compares its performance with two other CNNs trained separately on color and green channel fundus images. Illumination variation was addressed by correcting for the variability in the luminosity and contrast estimated from a large scale retinal regions. The CNN models were trained and evaluated on image patches extracted from a registered fundus image set collected from 51 diabetic eyes that were screened at two different time-points. The results show that using normalized images yield better performance than color and green channel images, suggesting that illumination normalization greatly facilitates CNNs to quickly and correctly learn distinctive local image features of DR related retinal changes.

Automated tissue characterization is one of the major applications of computer-aided diagnosis systems. Deep learning techniques have recently demonstrated impressive performance for the image patch-based tissue characterization. However, existing patch-based tissue classification techniques struggle to exploit the useful shape information. Local and global shape knowledge such as the regional boundary changes, diameter, and volumetrics can be useful in classifying the tissues especially in scenarios where the appearance signature does not provide significant classification information. In this work, we present a deep neural network-based method for the automated segmentation of the tumors referred to as optic pathway gliomas (OPG) located within the anterior visual pathway (AVP; optic nerve, chiasm or tracts) using joint shape and appearance learning. Voxel intensity values of commonly used MRI sequences are generally not indicative of OPG. To be considered an OPG, current clinical practice dictates that some portion of AVP must demonstrate shape enlargement. The method proposed in this work integrates multiple sequence magnetic resonance image (T1, T2, and FLAIR) along with local boundary changes to train a deep neural network. For training and evaluation purposes, we used a dataset of multiple sequence MRI obtained from 20 subjects (10 controls, 10 NF1+OPG). To our best knowledge, this is the first deep representation learning-based approach designed to merge shape and multi-channel appearance data for the glioma detection. In our experiments, mean misclassification errors of 2:39% and 0:48% were observed respectively for glioma and control patches extracted from the AVP. Moreover, an overall dice similarity coefficient of 0:87±0:13 (0:93±0:06 for healthy tissue, 0:78±0:18 for glioma tissue) demonstrates the potential of the proposed method in the accurate localization and early detection of OPG.

Deep learning and convolutional neural networks (CNNs) in particular are increasingly popular tools for segmentation and classification of medical images. CNNs were shown to be successful for segmentation of brain tumors into multiple regions or labels. However, in the environment which fosters data-sharing and collection of multi-institutional datasets, a question arises: does training with data from another institution with potentially different imaging equipment, contrast protocol, and patient population impact the segmentation performance of the CNN? Our study presents preliminary data towards answering this question. Specifically, we used MRI data of glioblastoma (GBM) patients for two institutions present in The Cancer Imaging Archive. We performed a process of training and testing CNN multiple times such that half of the time the CNN was tested on data from the same institution that was used for training and half of the time it was tested on another institution, keeping the training and testing set size constant. We observed a decrease in performance as measured by Dice coefficient when the CNN was trained with data from a different institution as compared to training with data from the same institution. The changes in performance for the entire tumor and for four different labels within the tumor were: 0.72 to 0.65 (p=0.06), 0.61 to 0.58 (p=0.49), 0.54 to 0.51 (p=0.82), 0.31 to 0.24 (p<0.03), and 0.43 to 0.31(p<0.003) respectively. In summary, we found that while data across institutions can be used for development of CNNs, this might be associated with a decrease in performance.

Tourette syndrome (TS) is a childhood-onset neurobehavioral disorder. To date, TS is still misdiagnosed due to its varied presentation and lacking of obvious clinical symptoms. Therefore, studies of objective imaging biomarkers are of great importance for early TS diagnosis. As tic generation has been linked to disturbed structural networks, and many efforts have been made recently to investigate brain functional or structural networks using machine learning methods, for the purpose of disease diagnosis. However, few studies were related to TS and some drawbacks still existed in them. Therefore, we propose a novel classification framework integrating a multi-threshold strategy and a network fusion scheme to address the preexisting drawbacks. Here we used diffusion MRI probabilistic tractography to construct the structural networks of 44 TS children and 48 healthy children. We ameliorated the similarity network fusion algorithm specially to fuse the multi-threshold structural networks. Graph theoretical analysis was then implemented, and nodal degree, nodal efficiency and nodal betweenness centrality were selected as features. Finally, support vector machine recursive feature extraction (SVM-RFE) algorithm was used for feature selection, and then optimal features are fed into SVM to automatically discriminate TS children from controls. We achieved a high accuracy of 89.13% evaluated by a nested cross validation, demonstrated the superior performance of our framework over other comparison methods. The involved discriminative regions for classification primarily located in the basal ganglia and frontal cortico-cortical networks, all highly related to the pathology of TS. Together, our study may provide potential neuroimaging biomarkers for early-stage TS diagnosis.

Multi-Voxel Magnetic Resonance Spectroscopy (MV-MRS) provides an important and insightful technique for the examination of the chemical composition of brain tissue, making it an attractive medical imaging modality for the examination of brain tumours. MRS, however, is affected by the issue of the Partial Volume Effect (PVE), where the signals of multiple tissue types can be found within a single voxel and provides an obstacle to the interpretation of the data. The PVE results from the low resolution achieved in MV-MRS images relating to the signal to noise ratio (SNR). To counteract PVE, this paper proposes a novel Pairwise Mixture Model (PMM), that extends a recently reported Signal Mixture Model (SMM) for representing the MV-MRS signal as normal, low or high grade tissue types. Inspired by Conditional Random Field (CRF) and its continuous variant the PMM incorporates the surrounding voxel neighbourhood into an optimisation problem, the solution of which provides an estimation to a set of coefficients. The values of the estimated coefficients represents the amount of each tissue type (normal, low or high) found within a voxel. These coefficients can then be visualised as a nosological rendering using a coloured grid representing the MV-MRS image overlaid on top of a structural image, such as a Magnetic Resonance Image (MRI). Experimental results show an accuracy of 92.69% in classifying patient tumours as either low or high grade compared against the histopathology for each patient. Compared to 91.96% achieved by the SMM, the proposed PMM method demonstrates the importance of incorporating spatial coherence into the estimation as well as its potential clinical usage.

Purpose: To 1) find effective texture features from multimodal MRI that can distinguish IDH mutant and wild status, and 2) propose a radiomic strategy for preoperatively detecting IDH mutation patients with glioma. Materials and Methods: 152 patients with glioma were retrospectively included from the Cancer Genome Atlas. Corresponding T1-weighted image before- and post-contrast, T2-weighted image and fluid-attenuation inversion recovery image from the Cancer Imaging Archive were analyzed. Specific statistical tests were applied to analyze the different kind of baseline information of LrGG patients. Finally, 168 texture features were derived from multimodal MRI per patient. Then the support vector machine-based recursive feature elimination (SVM-RFE) and classification strategy was adopted to find the optimal feature subset and build the identification models for detecting the IDH mutation. Results: Among 152 patients, 92 and 60 were confirmed to be IDH-wild and mutant, respectively. Statistical analysis showed that the patients without IDH mutation was significant older than patients with IDH mutation (p<0.01), and the distribution of some histological subtypes was significant different between IDH wild and mutant groups (p<0.01). After SVM-RFE, 15 optimal features were determined for IDH mutation detection. The accuracy, sensitivity, specificity, and AUC after SVM-RFE and parameter optimization were 82.2%, 85.0%, 78.3%, and 0.841, respectively. Conclusion: This study presented a radiomic strategy for noninvasively discriminating IDH mutation of patients with glioma. It effectively incorporated kinds of texture features from multimodal MRI, and SVM-based classification strategy. Results suggested that features selected from SVM-RFE were more potential to identifying IDH mutation. The proposed radiomics strategy could facilitate the clinical decision making in patients with glioma.

Recent studies showed that genomic analysis of lower grade gliomas can be very effective for stratification of patients into groups with different prognosis and proposed specific genomic classifications. In this study, we explore the association of one of those genomic classifications with imaging parameters to determine whether imaging could serve a similar role to genomics in cancer patient treatment. Specifically, we analyzed imaging and genomics data for 110 patients from 5 institutions from The Cancer Genome Atlas and The Cancer Imaging Archive datasets. The analyzed imaging data contained preoperative FLAIR sequence for each patient. The images were analyzed using the in-house algorithms which quantify 2D and 3D aspects of the tumor shape. Genomic data consisted of a cluster of clusters classification proposed in a very recent and leading publication in the field of lower grade glioma genomics. Our statistical analysis showed that there is a strong association between the tumor cluster-of-clusters subtype and two imaging features: bounding ellipsoid volume ratio and angular standard deviation. This result shows high promise for the potential use of imaging as a surrogate measure for genomics in the decision process regarding treatment of lower grade glioma patients.

Glioblastoma Multiforme (GBM) is a highly aggressive brain tumor with a median survival of 14 months. Hypoxia is a hallmark trait in GBM that is known to be associated with angiogenesis, tumor growth, and resistance to conventional therapy, thereby limiting treatment options for GBM patients. There is thus an urgent clinical need for non-invasively capturing tumor hypoxia in GBM towards identifying a subset of patients who would likely benefit from anti-angiogenic therapies (bevacizumab) in the adjuvant setting. In this study, we employed radiomic descriptors to (a) capture molecular variations of tumor hypoxia on routine MRI that are otherwise not appreciable; and (b) employ the radiomic correlates of hypoxia to discriminate patients with short-term survival (STS, overall survival (OS) < 7 months), mid-term survival (MTS) (7 months<OS<16 months), and long-term survival (LTS, OS>16 months). A total of 97 studies (25 STS, 36 MTS, 36 LTS) with Gadolinium T1-contrast (Gd-T1c), T2w, and FLAIR protocols with their corresponding gene expression profiles were obtained from the cancer genome atlas (TCGA) database. For each MRI study, necrotic, enhancing tumor, and edematous regions were segmented by an expert. A total of 30 radiomic descriptors (i.e. Haralick, Laws energy, Gabor) were extracted from every region across all three MRI protocols. By performing unsupervised clustering of the expression profile of hypoxia associated genes, a "low", "medium", or "high" index was defined for every study. Spearman correlation was then used to identify the most significantly correlated MRI features with the hypoxia index for every study. These features were further used to categorize each study as STS, MTS, and LTS using Kaplan-Meier (KM) analysis. Our results revealed that the most significant features (p < 0.05) were identified as Laws energy and Haralick features that capture image heterogeneity on FLAIR and Gd-T1w sequences. We also found these radiomic features to be significantly associated with survival, distinguishing MTS from LTS (p=.005) and STS from LTS (p=.0008).

With the recent success of machine learning based solutions for automatic image parsing, the availability of reference image annotations for algorithm training is one of the major bottlenecks in medical image segmentation. We are interested in interactive semantic segmentation methods that can be used in an online fashion to generate expert segmentations. These can be used to train automated segmentation techniques or, from an application perspective, for quick and accurate tumor progression monitoring.

Using simulated user interactions in a MRI glioblastoma segmentation task, we show that if the user possesses knowledge of the correct segmentation it is significantly (p ≤ 0.009) better to present data and current segmentation to the user in such a manner that they can easily identify falsely classified regions compared to guiding the user to regions where the classifier exhibits high uncertainty, resulting in differences of mean Dice scores between +0.070 (Whole tumor) and +0.136 (Tumor Core) after 20 iterations. The annotation process should cover all classes equally, which results in a significant (p ≤ 0.002) improvement compared to completely random annotations anywhere in falsely classified regions for small tumor regions such as the necrotic tumor core (mean Dice +0.151 after 20 it.) and non-enhancing abnormalities (mean Dice +0.069 after 20 it.). These findings provide important insights for the development of efficient interactive segmentation systems and user interfaces.

In dental X-ray images, an accurate detection of cephalometric landmarks plays an important role in clinical diagnosis, treatment and surgical decisions for dental problems. In this work, we propose an end-to-end deep learning system for cephalometric landmark detection in dental X-ray images, using convolutional neural networks (CNN). For detecting 19 cephalometric landmarks in dental X-ray images, we develop a detection system using CNN-based coordinate-wise regression systems. By viewing x- and y-coordinates of all landmarks as 38 independent variables, multiple CNN-based regression systems are constructed to predict the coordinate variables from input X-ray images. First, each coordinate variable is normalized by the length of either height or width of an image. For each normalized coordinate variable, a CNN-based regression system is trained on training images and corresponding coordinate variable, which is a variable to be regressed. We train 38 regression systems with the same CNN structure on coordinate variables, respectively. Finally, we compute 38 coordinate variables with these trained systems from unseen images and extract 19 landmarks by pairing the regressed coordinates. In experiments, the public database from the Grand Challenges in Dental X-ray Image Analysis in ISBI 2015 was used and the proposed system showed promising performance by successfully locating the cephalometric landmarks within considerable margins from the ground truths.

In this study, we propose an automated otoscopy image analysis system called Autoscope. To the best of our knowledge, Autoscope is the first system designed to detect a wide range of eardrum abnormalities by using high-resolution otoscope images and report the condition of the eardrum as “normal” or “abnormal.” In order to achieve this goal, first, we developed a preprocessing step to reduce camera-specific problems, detect the region of interest in the image, and prepare the image for further analysis. Subsequently, we designed a new set of clinically motivated eardrum features (CMEF). Furthermore, we evaluated the potential of the visual MPEG-7 descriptors for the task of tympanic membrane image classification. Then, we fused the information extracted from the CMEF and state-of-the-art computer vision features (CVF), which included MPEG-7 descriptors and two additional features together, using a state of the art classifier. In our experiments, 247 tympanic membrane images with 14 different types of abnormality were used, and Autoscope was able to classify the given tympanic membrane images as normal or abnormal with 84.6% accuracy.

The tympanic membrane (TM) is the bridging element between the pressure waves of sound in air and the ossicular chain. It allows for sound to be conducted into the inner ear, achieving the human sense of hearing. Otitis media with effusion (OME, commonly referred to as ‘glue ear’) is a typical condition in infants that prevents the vibration of the TM and causes conductive hearing loss, this can lead to stunting early stage development if undiagnosed. Furthermore, OME is hard to identify in this age group; as they cannot respond to typical audiometry tests. Tympanometry allows for the mobility of the TM to be examined without patient response, but requires expensive apparatus and specialist training. By combining a smartphone equipped with a 240 frames per second video recording capability with an otoscopic clip-on accessory, this paper presents a novel application of Eulerian Video Magnification (EVM) to video-otology, that could provide assistance in diagnosing OME. We present preliminary results showing a spatio-temporal slice taken from an exaggerated video visualization of the TM being excited in vivo on a healthy ear. Our preliminary results demonstrate the potential for using such an approach for diagnosing OME under visual inspection as alternative to tympanometry, which could be used remotely and hence help diagnosis in a wider population pool.

Accurate tumor progression prediction for oropharyngeal cancers is crucial for identifying patients who would best be treated with optimized treatment and therefore minimize the risk of under- or over-treatment. An objective decision support system that can merge the available radiomics, histopathologic and molecular biomarkers in a predictive model based on statistical outcomes of previous cases and machine learning may assist clinicians in making more accurate assessment of oropharyngeal tumor progression. In this study, we evaluated the feasibility of developing individual and combined predictive models based on quantitative image analysis from radiomics, histopathology and molecular biomarkers for oropharyngeal tumor progression prediction. With IRB approval, 31, 84, and 127 patients with head and neck CT (CT-HN), tumor tissue microarrays (TMAs) and molecular biomarker expressions, respectively, were collected. For 8 of the patients all 3 types of biomarkers were available and they were sequestered in a test set. The CT-HN lesions were automatically segmented using our level sets based method. Morphological, texture and molecular based features were extracted from CT-HN and TMA images, and selected features were merged by a neural network. The classification accuracy was quantified using the area under the ROC curve (AUC). Test AUCs of 0.87, 0.74, and 0.71 were obtained with the individual predictive models based on radiomics, histopathologic, and molecular features, respectively. Combining the radiomics and molecular models increased the test AUC to 0.90. Combining all 3 models increased the test AUC further to 0.94. This preliminary study demonstrates that the individual domains of biomarkers are useful and the integrated multi-domain approach is most promising for tumor progression prediction.

Segmentation of anatomical structures is fundamental in the development of computer aided diagnosis systems for cerebral pathologies. Manual annotations are laborious, time consuming and subject to human error and observer variability. Accurate quantification of cerebrospinal fluid (CSF) can be employed as a morphometric measure for diagnosis and patient outcome prediction. However, segmenting CSF in non-contrast CT images is complicated by low soft tissue contrast and image noise. In this paper we propose a state-of-the-art method using a multi-scale three-dimensional (3D) fully convolutional neural network (CNN) to automatically segment all CSF within the cranial cavity. The method is trained on a small dataset comprised of four manually annotated cerebral CT images. Quantitative evaluation of a separate test dataset of four images shows a mean Dice similarity coefficient of 0.87 ± 0.01 and mean absolute volume difference of 4.77 ± 2.70 %. The average prediction time was 68 seconds. Our method allows for fast and fully automated 3D segmentation of cerebral CSF in non-contrast CT, and shows promising results despite a limited amount of training data.

The imaging biomarkers EmphysemaPresence and NoduleSpiculation are crucial inputs for most models aiming to predict the risk of indeterminate pulmonary nodules detected at CT screening. To increase reproducibility and to accelerate screening workflow it is desirable to assess these biomarkers automatically. Validation on NLST images indicates that standard histogram measures are not sufficient to assess EmphysemaPresence in screenees. However, automatic scoring of bulla-resembling low attenuation areas can achieve agreement with experts with close to 80% sensitivity and specificity. NoduleSpiculation can be automatically assessed with similar accuracy. We find a dedicated spiculi tracing score to slightly outperform generic combinations of texture features with classifiers.

We investigated the feasibility of classifying pathologic invasive nodules and pre-invasive or benign nodules by quantitative analysis of the CT attenuation distribution patterns and other radiomic features of lung nodule components. We developed a new 3D adaptive multi-component Expectation-Maximization (EM) analysis method to segment the solid and non-solid nodule components and the surrounding lung parenchymal region. Features were extracted to characterize the size, shape, and the CT attenuation distribution of the entire nodule as well as the individual regions. With permission of the National Lung Screening Trial (NLST) project, a data set containing the baseline low dose CT scans of 53 cases with known pathologic tumor type categorization was obtained. The 53 cases contain 45 invasive nodules (group 1) and 42 pre-invasive nodules (group 2). A logistic regression model (LRM) was built using leave-one-case-out resampling and receiver operating characteristic (ROC) analysis for classification of group 1 and group 2, using the pathologic categorization as ground truth. With 4 selected features, the LRM achieved a test area under the curve (AUC) value of 0.877±0.036. The results demonstrated that the pathologic invasiveness of lung adenocarcinomas could be categorized according to the CT attenuation distribution patterns of the nodule components manifested on LDCT images.

Lung cancer is the most common cause of cancer-related death. To diagnose lung cancers in early stages, numerous studies and approaches have been developed for cancer screening with computed tomography (CT) imaging. In recent years, convolutional neural networks (CNN) have become one of the most common and reliable techniques in computer aided detection (CADe) and diagnosis (CADx) by achieving state-of-the-art-level performances for various tasks. In this study, we propose a CNN classification system for false positive reduction of initially detected lung nodule candidates. First, image patches of lung nodule candidates are extracted from CT scans to train a CNN classifier. To reflect the volumetric contextual information of lung nodules to 2D image patch, we propose a weighted average image patch (WAIP) generation by averaging multiple slice images of lung nodule candidates. Moreover, to emphasize central slices of lung nodules, slice images are locally weighted according to Gaussian distribution and averaged to generate the 2D WAIP. With these extracted patches, 2D CNN is trained to achieve the classification of WAIPs of lung nodule candidates into positive and negative labels. We used LUNA 2016 public challenge database to validate the performance of our approach for false positive reduction in lung CT nodule classification. Experiments show our approach improves the classification accuracy of lung nodules compared to the baseline 2D CNN with patches from single slice image.

In this paper, we investigated the problem of diagnostic lung nodule malignancy prediction using thoracic Computed Tomography (CT) screening. Unlike most existing studies classify the nodules into two types benign and malignancy, we interpreted the nodule malignancy prediction as a regression problem to predict continuous malignancy level. We proposed a joint multi-task learning algorithm using Convolutional Neural Network (CNN) to capture nodule heterogeneity by extracting discriminative features from alternatingly stacked layers. We trained a CNN regression model to predict the nodule malignancy, and designed a multi-task learning mechanism to simultaneously share knowledge among 9 different nodule characteristics (Subtlety, Calcification, Sphericity, Margin, Lobulation, Spiculation, Texture, Diameter and Malignancy), and improved the final prediction result. Each CNN would generate characteristic-specific feature representations, and then we applied multi-task learning on the features to predict the corresponding likelihood for that characteristic. We evaluated the proposed method on 2620 nodules CT scans from LIDC-IDRI dataset with the 5-fold cross validation strategy. The multitask CNN regression result for regression RMSE and mapped classification ACC were 0.830 and 83.03%, while the results for single task regression RMSE 0.894 and mapped classification ACC 74.9%. Experiments show that the proposed method could predict the lung nodule malignancy likelihood effectively and outperforms the state-of-the-art methods. The learning framework could easily be applied in other anomaly likelihood prediction problem, such as skin cancer and breast cancer. It demonstrated the possibility of our method facilitating the radiologists for nodule staging assessment and individual therapeutic planning.

To determine the potential usefulness of quantified diagnostic image features as inputs to a CAD system, we investigate the predictive capabilities of statistical learning methods for classifying nodule malignancy, utilizing the Lung Image Database Consortium (LIDC) dataset, and only employ the radiologist-assigned diagnostic feature values for the lung nodules therein, as well as our derived estimates of the diameter and volume of the nodules from the radiologists' annotations. We calculate theoretical upper bounds on the classification accuracy that is achievable by an ideal classifier that only uses the radiologist-assigned feature values, and we obtain an accuracy of 85.74 (±1.14)% which is, on average, 4.43% below the theoretical maximum of 90.17%. The corresponding area-under-the-curve (AUC) score is 0.932 (±0.012), which increases to 0.949 (±0.007) when diameter and volume features are included, along with the accuracy to 88.08 (±1.11)%. Our results are comparable to those in the literature that use algorithmically-derived image-based features, which supports our hypothesis that lung nodules can be classified as malignant or benign using only quantified, diagnostic image features, and indicates the competitiveness of this approach. We also analyze how the classification accuracy depends on specific features, and feature subsets, and we rank the features according to their predictive power, statistically demonstrating the top four to be spiculation, lobulation, subtlety, and calcification.

Patient-targeted treatment of non-small cell lung carcinoma (NSCLC) has been well documented according to the histologic subtypes over the past decade. In parallel, recent development of quantitative image biomarkers has recently been highlighted as important diagnostic tools to facilitate histological subtype classification. In this study, we present a radiomics analysis that classifies the adenocarcinoma (ADC) and squamous cell carcinoma (SqCC). We extract 52-dimensional, CT-based features (7 statistical features and 45 image texture features) to represent each nodule. We evaluate our approach on a clinical dataset including 324 ADCs and 110 SqCCs patients with CT image scans. Classification of these features is performed with four different machine-learning classifiers including Support Vector Machines with Radial Basis Function kernel (RBF-SVM), Random forest (RF), K-nearest neighbor (KNN), and RUSBoost algorithms. To improve the classifiers’ performance, optimal feature subset is selected from the original feature set by using an iterative forward inclusion and backward eliminating algorithm. Extensive experimental results demonstrate that radiomics features achieve encouraging classification results on both complete feature set (AUC=0.89) and optimal feature subset (AUC=0.91).

Support vector machines (SVM) are applied to the problem of prostate lesion classification for the SPIE ProstateX Challenge 2016, achieving a score of 0.82 AUC on held-out test data. Square 5mm transverse image patches are extracted around each lesion center from aligned MRI scans. Three MRI modalities are simultaneously analyzed: T2-weighted, apparent diffusion coefficient (ADC) and volume transfer constant (K^trans). Extracted patches are used to train a binary classifier to predict clinical significance. The machine learning algorithm is trained on 76 positive cases and 254 negative cases (330 total) from the challenge. The method is conceptually simple, trains in a few seconds and yields competitive results.

A novel deep learning architecture (XmasNet) based on convolutional neural networks was developed for the classification of prostate cancer lesions, using the 3D multiparametric MRI data provided by the PROSTATEx challenge. End-to-end training was performed for XmasNet, with data augmentation done through 3D rotation and slicing, in order to incorporate the 3D information of the lesion. XmasNet outperformed traditional machine learning models based on engineered features, for both train and test data. For the test data, XmasNet outperformed 69 methods from 33 participating groups and achieved the second highest AUC (0.84) in the PROSTATEx challenge. This study shows the great potential of deep learning for cancer imaging.

In this manuscript, we describe our approach and methods to the ProstateX challenge, which achieved an overall AUC of 0.84 and the runner-up position. We train a deep convolutional neural network to classify lesions marked on multiparametric MRI of the prostate as clinically significant or not.

We implement a novel addition to the standard convolutional architecture described as auto-windowing which is clinically inspired and designed to overcome some of the difficulties faced in MRI interpretation, where high dynamic ranges and low contrast edges may cause difficulty for traditional convolutional neural networks trained on high contrast natural imagery. We demonstrate that this system can be trained end to end and outperforms a similar architecture without such additions. Although a relatively small training set was provided, we use extensive data augmentation to prevent overfitting and transfer learning to improve convergence speed, showing that deep convolutional neural networks can be feasibly trained on small datasets.

Prostate cancer (PCa) remains a leading cause of cancer mortality among American men. Multi-parametric magnetic resonance imaging (mpMRI) is widely used to assist with detection of PCa and characterization of its aggressiveness. Computer-aided diagnosis (CADx) of PCa in MRI can be used as clinical decision support system to aid radiologists in interpretation and reporting of mpMRI. We report on the development of a convolution neural network (CNN) model to support CADx in PCa based on the appearance of prostate tissue in mpMRI, conducted as part of the SPIE-AAPM-NCI PROSTATEx challenge. The performance of different combinations of mpMRI inputs to CNN was assessed and the best result was achieved using DWI and DCE-MRI modalities together with the zonal information of the finding. On the test set, the model achieved an area under the receiver operating characteristic curve of 0.80.

¹⁸F-DMFP-PET is a neuroimaging modality that allows us to analyze the striatal dopamine. Thus, it is recently emerging as an effective tool to assist the diagnosis of Parkinsonism and differentiate among parkinsonian syndromes. However the analysis of these data, which require specific preprocessing methods, is still poorly covered. In this work we demonstrate a novel methodology based on Hidden Markov Random Fields (HMRF) and the Gaussian distribution to preprocess ¹⁸F-DMFP-PET data. First, we performed a selection of voxels based on the analysis of the histogram in order to remove low-signal regions and regions outside the brain. Specifically, we modeled the histogram of intensities of a neuroimage with a mixture of two Gaussians and then, using a HMRF algorithm the voxels corresponding to the low-intensity Gaussian were discarded. This procedure is similar to the tissue segmentation usually applied to Magnetic Resonance Imaging data. Secondly, the intensity of the selected voxels was scaled so that the Gaussian that models the histogram for each neuroimage has same mean and standard deviation. This step made comparable the data from different patients, without removing the characteristic patterns of each patient's disorder. The proposed approach was evaluated using a computer system based on statistical classification that separated the neuroimages according to the parkinsonian variant they represented. The proposed approach achieved higher accuracy rates than standard approaches for voxel selection (based on atlases) and intensity normalization (based on the global mean).

The early diagnosis of idiopathic normal pressure hydrocephalus (iNPH) considered as a treatable dementia is important. The iNPH causes enlargement of lateral ventricles (LVs). The degree of the enlargement of the LVs on CT or MR images is evaluated by using a diagnostic imaging criterion, Evans index. Evans index is defined as the ratio of the maximal width of frontal horns (FH) of the LVs to the maximal width of the inner skull (IS). Evans index is the most commonly used parameter for the evaluation of ventricular enlargement. However, manual measurement of Evans index is a time-consuming process. In this study, we present an automated method to compute Evans index on brain CT images. The algorithm of the method consisted of five major steps: standardization of CT data to an atlas, extraction of FH and IS regions, the search for the outmost points of bilateral FH regions, determination of the maximal widths of both the FH and the IS, and calculation of Evans index. The standardization to the atlas was performed by using linear affine transformation and non-linear wrapping techniques. The FH regions were segmented by using a three dimensional region growing technique. This scheme was applied to CT scans from 44 subjects, including 13 iNPH patients. The average difference in Evans index between the proposed method and manual measurement was 0.01 (1.6%), and the correlation coefficient of these data for the Evans index was 0.98. Therefore, this computerized method may have the potential to accurately compute Evans index for the diagnosis of iNPH on CT images.

Statistical learning and decision theory play a key role in many areas of science and engineering. Some examples include time series regression and prediction, optical character recognition, signal detection in communications or biomedical applications for diagnosis and prognosis. This paper deals with the topic of learning from biomedical image data in the classification problem. In a typical scenario we have a training set that is employed to fit a prediction model or learner and a testing set on which the learner is applied to in order to predict the outcome for new unseen patterns. Both processes are usually completely separated to avoid over-fitting and due to the fact that, in practice, the unseen new objects (testing set) have unknown outcomes. However, the outcome yields one of a discrete set of values, i.e. the binary diagnosis problem. Thus, assumptions on these outcome values could be established to obtain the most likely prediction model at the training stage, that could improve the overall classification accuracy on the testing set, or keep its performance at least at the level of the selected statistical classifier. In this sense, a novel case-based learning (c-learning) procedure is proposed which combines hypothesis testing from a discrete set of expected outcomes and a cross-validated classification stage.

Prediction of survival time from brain tumor magnetic resonance images (MRI) is not commonly performed and would ordinarily be a time consuming process. However, current cross-sectional imaging techniques, particularly MRI, can be used to generate many features that may provide information on the patient’s prognosis, including survival. This information can potentially be used to identify individuals who would benefit from more aggressive therapy. Rather than using pre-defined and hand-engineered features as with current radiomics methods, we investigated the use of deep features extracted from pre-trained convolutional neural networks (CNNs) in predicting survival time. We also provide evidence for the power of domain specific fine-tuning in improving the performance of a pre-trained CNN’s, even though our data set is small. We fine-tuned a CNN initially trained on a large natural image recognition dataset (Imagenet ILSVRC) and transferred the learned feature representations to the survival time prediction task, obtaining over 81% accuracy in a leave one out cross validation.

The world’s aging population has given rise to an increasing awareness towards neurodegenerative disorders, including Alzheimers Disease (AD). Treatment options for AD are currently limited, but it is believed that future success depends on our ability to detect the onset of the disease in its early stages. The most frequently used tools for this include neuropsychological assessments, along with genetic, proteomic, and image-based diagnosis. Recently, the applicability of Diffusion Magnetic Resonance Imaging (dMRI) analysis for early diagnosis of AD has also been reported. The sensitivity of dMRI to the microstructural organization of cerebral tissue makes it particularly well-suited to detecting changes which are known to occur in the early stages of AD. Existing dMRI approaches can be divided into two broad categories: region-based and tract-based. In this work, we propose a new approach, which extends region-based approaches to the simultaneous characterization of multiple brain regions. Given a predefined set of features derived from dMRI data, we compute the probabilistic distances between different brain regions and treat the resulting connectivity pattern as an undirected, fully-connected graph. The characteristics of this graph are then used as markers to discriminate between AD subjects and normal controls (NC). Although in this preliminary work we omit subjects in the prodromal stage of AD, mild cognitive impairment (MCI), our method demonstrates perfect separability between AD and NC subject groups with substantial margin, and thus holds promise for fine-grained stratification of NC, MCI and AD populations.

Traumatic brain injury (TBI) is a major cause of death and disability in the United States. Time to treatment is often related to patient outcome. Access to cerebral imaging data in a timely manner is a vital component of patient care. Current methods of detecting and quantifying intracranial pathology can be time-consuming and require careful review of 2D/3D patient images by a radiologist. Additional time is needed for image protocoling, acquisition, and processing. These steps often occur in series, adding more time to the process and potentially delaying time-dependent management decisions for patients with traumatic brain injury.

Our team adapted machine learning and computer vision methods to develop a technique that rapidly and automatically detects CT-identifiable lesions. Specifically, we use scale invariant feature transform (SIFT)¹ and deep convolutional neural networks (CNN)² to identify important image features that can distinguish TBI lesions from background data. Our learning algorithm is a linear support vector machine (SVM)³. Further, we also employ tools from topological data analysis (TDA) for gleaning insights into the correlation patterns between healthy and pathological data. The technique was validated using 409 CT scans of the brain, acquired via the Progesterone for the Treatment of Traumatic Brain Injury phase III clinical trial (ProTECT_III) which studied patients with moderate to severe TBI⁴. CT data were annotated by a central radiologist and included patients with positive and negative scans. Additionally, the largest lesion on each positive scan was manually segmented. We reserved 80% of the data for training the SVM and used the remaining 20% for testing. Preliminary results are promising with 92.55% prediction accuracy (sensitivity = 91.15%, specificity = 93.45%), indicating the potential usefulness of this technique in clinical scenarios.

Parkinsonian syndromes encompass a spectrum of neurodegenerative diseases, which can be classified into various subtypes. The differentiation of these subtypes is typically conducted based on clinical criteria. Due to the overlap of intra-syndrome symptoms, the accurate differential diagnosis based on clinical guidelines remains a challenge with failure rates up to 25%. The aim of this study is to present an image-based classification method of patients with Parkinson’s disease (PD) and patients with progressive supranuclear palsy (PSP), an atypical variant of PD. Therefore, apparent diffusion coefficient (ADC) parameter maps were calculated based on diffusion-tensor magnetic resonance imaging (MRI) datasets. Mean ADC values were determined in 82 brain regions using an atlas-based approach. The extracted mean ADC values for each patient were then used as features for classification using a linear kernel support vector machine classifier. To increase the classification accuracy, a feature selection was performed, which resulted in the top 17 attributes to be used as the final input features. A leave-one-out cross validation based on 56 PD and 21 PSP subjects revealed that the proposed method is capable of differentiating PD and PSP patients with an accuracy of 94.8%. In conclusion, the classification of PD and PSP patients based on ADC features obtained from diffusion MRI datasets is a promising new approach for the differentiation of Parkinsonian syndromes in the broader context of decision support systems.

Detection of residual brain tumor is important to evaluate efficacy of brain cancer surgery, determine optimal strategy of further radiation therapy if needed, and assess ultimate prognosis of the patients. Brain MR is a commonly used imaging modality for this task. In order to distinguish between residual tumor and surgery induced scar tissues, two sets of MRI scans are conducted pre- and post-gadolinium contrast injection. The residual tumors are only enhanced in the post-contrast injection images. However, subjective reading and quantifying this type of brain MR images faces difficulty in detecting real residual tumor regions and measuring total volume of the residual tumor. In order to help solve this clinical difficulty, we developed and tested a new interactive computer-aided detection scheme, which consists of three consecutive image processing steps namely, 1) segmentation of the intracranial region, 2) image registration and subtraction, 3) tumor segmentation and refinement. The scheme also includes a specially designed and implemented graphical user interface (GUI) platform. When using this scheme, two sets of pre- and post-contrast injection images are first automatically processed to detect and quantify residual tumor volume. Then, a user can visually examine segmentation results and conveniently guide the scheme to correct any detection or segmentation errors if needed. The scheme has been repeatedly tested using five cases. Due to the observed high performance and robustness of the testing results, the scheme is currently ready for conducting clinical studies and helping clinicians investigate the association between this quantitative image marker and outcome of patients.

Confocal laser endomicroscopy (CLE), although capable of obtaining images at cellular resolution during surgery of brain tumors in real time, creates as many non-diagnostic as diagnostic images. Non-useful images are often distorted due to relative motion between probe and brain or blood artifacts. Many images, however, simply lack diagnostic features immediately informative to the physician. Examining all the hundreds or thousands of images from a single case to discriminate diagnostic images from nondiagnostic ones can be tedious. Providing a real time diagnostic value assessment of images (fast enough to be used during the surgical acquisition process and accurate enough for the pathologist to rely on) to automatically detect diagnostic frames would streamline the analysis of images and filter useful images for the pathologist/surgeon. We sought to automatically classify images as diagnostic or non-diagnostic. AlexNet, a deep-learning architecture, was used in a 4-fold cross validation manner. Our dataset includes 16,795 images (8572 nondiagnostic and 8223 diagnostic) from 74 CLE-aided brain tumor surgery patients. The ground truth for all the images is provided by the pathologist. Average model accuracy on test data was 91% overall (90.79 % accuracy, 90.94 % sensitivity and 90.87 % specificity). To evaluate the model reliability we also performed receiver operating characteristic (ROC) analysis yielding 0.958 average for area under ROC curve (AUC). These results demonstrate that a deeply trained AlexNet network can achieve a model that reliably and quickly recognizes diagnostic CLE images.

Stroke is a leading cause of death and disability in the western hemisphere. Acute ischemic strokes can be broadly classified based on the underlying cause into atherosclerotic strokes, cardioembolic strokes, small vessels disease, and stroke with other causes. The ability to determine the exact origin of an acute ischemic stroke is highly relevant for optimal treatment decision and preventing recurrent events. However, the differentiation of atherosclerotic and cardioembolic phenotypes can be especially challenging due to similar appearance and symptoms. The aim of this study was to develop and evaluate the feasibility of an image-based machine learning approach for discriminating between arteriosclerotic and cardioembolic acute ischemic strokes using 56 apparent diffusion coefficient (ADC) datasets from acute stroke patients. For this purpose, acute infarct lesions were semi-atomically segmented and 30,981 geometric and texture image features were extracted for each stroke volume. To improve the performance and accuracy, categorical Pearson’s χ² test was used to select the most informative features while removing redundant attributes. As a result, only 289 features were finally included for training of a deep multilayer feed-forward neural network without bootstrapping. The proposed method was evaluated using a leave-one-out cross validation scheme. The proposed classification method achieved an average area under receiver operator characteristic curve value of 0.93 and a classification accuracy of 94.64%. These first results suggest that the proposed image-based classification framework can support neurologists in clinical routine differentiating between atherosclerotic and cardioembolic phenotypes.

In this work, we propose a novel method to improve texture based tumor segmentation by fusing cell density patterns that are generated from tumor growth modeling. To model tumor growth, we solve the reaction-diffusion equation by using Lattice-Boltzmann method (LBM). Computational tumor growth modeling obtains the cell density distribution that potentially indicates the predicted tissue locations in the brain over time. The density patterns is then considered as novel features along with other texture (such as fractal, and multifractal Brownian motion (mBm)), and intensity features in MRI for improved brain tumor segmentation. We evaluate the proposed method with about one hundred longitudinal MRI scans from five patients obtained from public BRATS 2015 data set, validated by the ground truth. The result shows significant improvement of complete tumor segmentation using ANOVA analysis for five patients in longitudinal MR images.

Breast cancer is the leading cause of death for women in most countries. The high levels of mortality relate mostly to late diagnosis and to the direct proportionally relationship between breast density and breast cancer development. Therefore, the correct assessment of breast density is important to provide better screening for higher risk patients. However, in modern digital mammography the discrimination among breast densities is highly complex due to increased contrast and visual information for all densities. Thus, a computational system for classifying breast density might be a useful tool for aiding medical staff. Several machine-learning algorithms are already capable of classifying small number of classes with good accuracy. However, machinelearning algorithms main constraint relates to the set of features extracted and used for classification. Although well-known feature extraction techniques might provide a good set of features, it is a complex task to select an initial set during design of a classifier. Thus, we propose feature extraction using a Convolutional Neural Network (CNN) for classifying breast density by a usual machine-learning classifier. We used 307 mammographic images downsampled to 260x200 pixels to train a CNN and extract features from a deep layer. After training, the activation of 8 neurons from a deep fully connected layer are extracted and used as features. Then, these features are feedforward to a single hidden layer neural network that is cross-validated using 10-folds to classify among four classes of breast density. The global accuracy of this method is 98.4%, presenting only 1.6% of misclassification. However, the small set of samples and memory constraints required the reuse of data in both CNN and MLP-NN, therefore overfitting might have influenced the results even though we cross-validated the network. Thus, although we presented a promising method for extracting features and classifying breast density, a greater database is still required for evaluating the results.

Mammography analysis is an important tool that helps detecting breast cancer at the very early stages of the disease, thus increasing the quality of life of hundreds of thousands of patients worldwide. In Computer-Aided Detection systems, the identification of mammograms with and without masses (without clinical findings) is highly needed to reduce the false positive rates regarding the automatic selection of regions of interest that may contain some suspicious content. In this work, the introduce a variant of the Optimum-Path Forest (OPF) classifier for breast mass identification, as well as we employed an ensemble-based approach that can enhance the effectiveness of individual classifiers aiming at dealing with the aforementioned purpose. The experimental results also comprise the naïve OPF and a traditional neural network, being the most accurate results obtained through the ensemble of classifiers, with an accuracy nearly to 86%.

Mammographic breast density is a well-established marker for breast cancer risk. However, accurate measurement of dense tissue is a difficult task due to faint contrast and significant variations in background fatty tissue. This study presents a novel method for automated mammographic density estimation based on Convolutional Neural Network (CNN). A total of 397 full-field digital mammograms were selected from Seoul National University Hospital. Among them, 297 mammograms were randomly selected as a training set and the rest 100 mammograms were used for a test set. We designed a CNN architecture suitable to learn the imaging characteristic from a multitudes of sub-images and classify them into dense and fatty tissues. To train the CNN, not only local statistics but also global statistics extracted from an image set were used. The image set was composed of original mammogram and eigen-image which was able to capture the X-ray characteristics in despite of the fact that CNN is well known to effectively extract features on original image. The 100 test images which was not used in training the CNN was used to validate the performance. The correlation coefficient between the breast estimates by the CNN and those by the expert’s manual measurement was 0.96. Our study demonstrated the feasibility of incorporating the deep learning technology into radiology practice, especially for breast density estimation. The proposed method has a potential to be used as an automated and quantitative assessment tool for mammographic breast density in routine practice.

High mammographic density is associated with an increased risk of breast cancer, however whether the association is stronger when there is agreement across measures is unclear. This study investigates whether a combination of density measures is a better predictor of breast cancer risk than individual methods alone. Women recruited to the Predicting Risk of Cancer At Screening (PROCAS) study and with mammographic density assessed using three different methods were included (n=33,304). Density was assessed visually using Visual Analogue Scales (VAS) and by two fully automated methods, Quantra and Volpara. Percentage breast density was divided into (high, medium and low) and combinations of measures were used to further categorise individuals (e.g. ‘all high’). A total of 667 breast cancers were identified and logistic regression was used to determine the relationship between breast density and breast cancer risk. In total, 44% of individuals were in the same tertile for all three measures, 8.6% were in non-adjacent (high and low) or mixed categories (high, medium and low). For individual methods the strongest association with breast cancer risk was for medium and high tertiles of VAS with odds ratios (OR) adjusted for age and BMI of 1.63 (95% CI 1.31-2.03) and 2.33 (1.87-2.90) respectively. For the combination of density methods the strongest association was for ‘all high’ (OR 2.42, 1.77-3.31) followed by “two high” (OR 1.90, 1.35-3.31) and “two medium” (OR 1.88, 1.40-2.52). Combining density measures did not affect the magnitude of risk compared to using individual methods.

This paper describes a novel approach for the automatic assessment of breast density in non-contrast three-dimensional computed tomography (3D CT) images. The proposed approach trains and uses a deep convolutional neural network (CNN) from scratch to classify breast tissue density directly from CT images without segmenting the anatomical structures, which creates a bottleneck in conventional approaches. Our scheme determines breast density in a 3D breast region by decomposing the 3D region into several radial 2D-sections from the nipple, and measuring the distribution of breast tissue densities on each 2D section from different orientations. The whole scheme is designed as a compact network without the need for post-processing and provides high robustness and computational efficiency in clinical settings. We applied this scheme to a dataset of 463 non-contrast CT scans obtained from 30- to 45-year-old-women in Japan. The density of breast tissue in each CT scan was assigned to one of four categories (glandular tissue within the breast <25%, 25%–50%, 50%–75%, and >75%) by a radiologist as ground truth. We used 405 CT scans for training a deep CNN and the remaining 58 CT scans for testing the performance. The experimental results demonstrated that the findings of the proposed approach and those of the radiologist were the same in 72% of the CT scans among the training samples and 76% among the testing samples. These results demonstrate the potential use of deep CNN for assessing breast tissue density in non-contrast 3D CT images.

Mammography is the most efficient modality for detection of breast cancer at early stage. Microcalcifications are tiny bright spots in mammograms and can often get missed by the radiologist during diagnosis. The presence of microcalcification clusters in mammograms can act as an early sign of breast cancer. This paper presents a completely automated computer-aided detection (CAD) system for detection of microcalcification clusters in mammograms. Unsharp masking is used as a preprocessing step which enhances the contrast between microcalcifications and the background. The preprocessed image is thresholded and various shape and intensity based features are extracted. Support vector machine (SVM) classifier is used to reduce the false positives while preserving the true microcalcification clusters. The proposed technique is applied on two different databases i.e DDSM and private database. The proposed technique shows good sensitivity with moderate false positives (FPs) per image on both databases.

Breast cancer screening with mammography and ultrasonography is expected to improve sensitivity compared with mammography alone, especially for women with dense breast. An automated breast volume scanner (ABVS) provides the operator-independent whole breast data which facilitate double reading and comparison with past exams, contralateral breast, and multimodality images. However, large volumetric data in screening practice increase radiologists’ workload. Therefore, our goal is to develop a computer-aided detection scheme of breast masses in ABVS data for assisting radiologists’ diagnosis and comparison with mammographic findings. In this study, false positive (FP) reduction scheme using deep convolutional neural network (DCNN) was investigated. For training DCNN, true positive and FP samples were obtained from the result of our initial mass detection scheme using the vector convergence filter. Regions of interest including the detected regions were extracted from the multiplanar reconstraction slices. We investigated methods to select effective FP samples for training the DCNN. Based on the free response receiver operating characteristic analysis, simple random sampling from the entire candidates was most effective in this study. Using DCNN, the number of FPs could be reduced by 60%, while retaining 90% of true masses. The result indicates the potential usefulness of DCNN for FP reduction in automated mass detection on ABVS images.

Breast elastography is a new sonographic technique that provides additional information to evaluate tissue stiffness. However, interpreting breast elastography images can vary depending on the radiologist. In order to provide quantitative and less subjective data regarding the stiffness of a lesion, we developed a tool to measure the amount of hard area in a lesion from the 2D image. The database consisted of 78 patients with 83 breast lesions (31 malignant and 52 benign). Two radiologists and one resident manually drew the contour of the lesions in B-mode ultrasound images and the contour was mapped in the elastography image. By using the system proposed, the radiologists obtained a very good diagnostic agreement among themselves (kappa = 0.86), achieving the same sensitivity and specificity (80.7 and 88.5, respectively), and an AUC of 0.883 for Radiologist 1 and 0.892 for Radiologist 2. The Resident had less interobserver agreement, as well as lower specificity and AUC, which may be related to less experience. Furthermore, the radiologists had an agreement with the tool used in the automatic method higher than 90%. Thus, the method developed was useful in aiding the diagnosis of breast lesions in strain elastography, minimizing its subjectivity.

This paper presents a new local micro-pattern texture descriptor for the detection of Architectural Distortion (AD) in digital mammography images. AD is a subtle contraction of breast parenchyma that may represent an early sign of breast cancer. Due to its subtlety and variability, AD is more difficult to detect compared to microcalcifications and masses, and is commonly found in retrospective evaluations of false-negative mammograms. Several computer-based systems have been proposed for automatic detection of AD, but their performance are still unsatisfactory. The proposed descriptor, Local Mapped Pattern (LMP), is a generalization of the Local Binary Pattern (LBP), which is considered one of the most powerful feature descriptor for texture classification in digital images. Compared to LBP, the LMP descriptor captures more effectively the minor differences between the local image pixels. Moreover, LMP is a parametric model which can be optimized for the desired application. In our work, the LMP performance was compared to the LBP and four Haralick's texture descriptors for the classification of 400 regions of interest (ROIs) extracted from clinical mammograms. ROIs were selected and divided into four classes: AD, normal tissue, microcalcifications and masses. Feature vectors were used as input to a multilayer perceptron neural network, with a single hidden layer. Results showed that LMP is a good descriptor to distinguish AD from other anomalies in digital mammography. LMP performance was slightly better than the LBP and comparable to Haralick's descriptors (mean classification accuracy = 83%).

Personalised breast screening requires assessment of individual risk of breast cancer, of which one contributory factor is weight. Self-reported weight has been used for this purpose, but may be unreliable. We explore the use of volume of fat in the breast, measured from digital mammograms. Volumetric breast density measurements were used to determine the volume of fat in the breasts of 40,431 women taking part in the Predicting Risk Of Cancer At Screening (PROCAS) study. Tyrer-Cuzick risk using self-reported weight was calculated for each woman. Weight was also estimated from the relationship between self-reported weight and breast fat volume in the cohort, and used to re-calculate Tyrer-Cuzick risk. Women were assigned to risk categories according to 10 year risk (below average <2%, average 2-3.49%, above average 3.5-4.99%, moderate 5-7.99%, high ≥8%) and the original and re-calculated Tyrer-Cuzick risks were compared. Of the 716 women diagnosed with breast cancer during the study, 15 (2.1%) moved into a lower risk category, and 37 (5.2%) moved into a higher category when using weight estimated from breast fat volume. Of the 39,715 women without a cancer diagnosis, 1009 (2.5%) moved into a lower risk category, and 1721 (4.3%) into a higher risk category. The majority of changes were between below average and average risk categories (38.5% of those with a cancer diagnosis, and 34.6% of those without). No individual moved more than one risk group. Automated breast fat measures may provide a suitable alternative to self-reported weight for risk assessment in personalized screening.

Computer-Aided Diagnosis of masses in mammograms is an important indicator of breast cancer. The use of retrieval systems in breast examination is increasing gradually. In this respect, the method of exploiting the vocabulary tree framework and the inverted file in the mammographic masse retrieval have been proved high accuracy and excellent scalability. However it just considered the features in each image as a visual word and had ignored the spatial configurations of features. It greatly affect the retrieval performance. To overcome this drawback, we introduce the geometric verification method to retrieval in mammographic masses. First of all, we obtain corresponding match features based on the vocabulary tree framework and the inverted file. After that, we grasps the main point of local similarity characteristic of deformations in the local regions by constructing the circle regions of corresponding pairs. Meanwhile we segment the circle to express the geometric relationship of local matches in the area and generate the spatial encoding strictly. Finally we judge whether the matched features are correct or not, based on verifying the all spatial encoding are whether satisfied the geometric consistency. Experiments show the promising results of our approach.

Reducing the overdiagnosis and overtreatment associated with ductal carcinoma in situ (DCIS) requires accurate prediction of the invasive potential at cancer screening. In this work, we investigated the utility of pre-operative histologic and mammographic features to predict upstaging of DCIS. The goal was to provide intentionally conservative baseline performance using readily available data from radiologists and pathologists and only linear models. We conducted a retrospective analysis on 99 patients with DCIS. Of those 25 were upstaged to invasive cancer at the time of definitive surgery. Pre-operative factors including both the histologic features extracted from stereotactic core needle biopsy (SCNB) reports and the mammographic features annotated by an expert breast radiologist were investigated with statistical analysis. Furthermore, we built classification models based on those features in an attempt to predict the presence of an occult invasive component in DCIS, with generalization performance assessed by receiver operating characteristic (ROC) curve analysis. Histologic features including nuclear grade and DCIS subtype did not show statistically significant differences between cases with pure DCIS and with DCIS plus invasive disease. However, three mammographic features, i.e., the major axis length of DCIS lesion, the BI-RADS level of suspicion, and radiologist’s assessment did achieve the statistical significance. Using those three statistically significant features as input, a linear discriminant model was able to distinguish patients with DCIS plus invasive disease from those with pure DCIS, with AUC-ROC equal to 0.62. Overall, mammograms used for breast screening contain useful information that can be perceived by radiologists and help predict occult invasive components in DCIS.

According to the World Health Organization, cardiovascular diseases are the leading cause of death worldwide, accounting for 17.3 million deaths per year, a number that is expected to grow to more than 23.6 million by 2030. Most cardiac pathologies involve the left ventricle; therefore, estimation of several functional parameters from a previous segmentation of this structure can be helpful in diagnosis. Manual delineation is a time consuming and tedious task that is also prone to high intra and inter-observer variability. Thus, there exists a need for automated cardiac segmentation method to help facilitate the diagnosis of cardiovascular diseases. In this work we propose a deep fully convolutional neural network architecture to address this issue and assess its performance. The model was trained end to end in a supervised learning stage from whole cardiac MRI images input and ground truth to make a per pixel classification. For its design, development and experimentation was used Caffe deep learning framework over an NVidia Quadro K4200 Graphics Processing Unit. The net architecture is: Conv64-ReLU (2x) – MaxPooling – Conv128-ReLU (2x) – MaxPooling – Conv256-ReLU (2x) – MaxPooling – Conv512-ReLu-Dropout (2x) – Conv2-ReLU – Deconv – Crop – Softmax. Training and testing processes were carried out using 5-fold cross validation with short axis cardiac magnetic resonance images from Sunnybrook Database. We obtained a Dice score of 0.92 and 0.90, Hausdorff distance of 4.48 and 5.43, Jaccard index of 0.97 and 0.97, sensitivity of 0.92 and 0.90 and specificity of 0.99 and 0.99, overall mean values with SGD and RMSProp, respectively.

Content-based image retrieval (CBIR) aims at retrieving from a database objects that are similar to an object provided by a query, by taking into consideration a set of extracted features. While CBIR has been widely applied in the two-dimensional image domain, the retrieval of3D objects from medical image datasets using CBIR remains to be explored. In this context, the development of descriptors that can capture information specific to organs or structures is desirable. In this work, we focus on the retrieval of two anatomical structures commonly imaged by Magnetic Resonance Imaging (MRI) and Computed Tomography (CT) techniques, the left ventricle of the heart and blood vessels. Towards this aim, we developed the Area-Distance Local Descriptor (ADLD), a novel 3D local shape descriptor that employs mesh geometry information, namely facet area and distance from centroid to surface, to identify shape changes. Because ADLD only considers surface meshes extracted from volumetric medical images, it substantially diminishes the amount of data to be analyzed. A 90% precision rate was obtained when retrieving both convex (left ventricle) and non-convex structures (blood vessels), allowing for detection of abnormalities associated with changes in shape. Thus, ADLD has the potential to aid in the diagnosis of a wide range of vascular and cardiac diseases.

In this paper we describe a pilot study, where machine learning methods are used to differentiate between congenital heart diseases. Our approach was to apply convolutional neural networks (CNNs) to echocardiographic images from five different pediatric populations: normal, coarctation of the aorta (CoA), hypoplastic left heart syndrome (HLHS), transposition of the great arteries (TGA), and single ventricle (SV). We used a single network topology that was trained in a pairwise fashion in order to evaluate the potential to differentiate between patient populations. In total we used 59,151 echo frames drawn from 1,666 clinical sequences. Approximately 80% of the data was used for training, and the remainder for validation. Data was split at sequence boundaries to avoid having related images in the training and validation sets. While training was done with echo images/frames, evaluation was performed for both single frame discrimination as well as sequence discrimination (by majority voting). In total 10 networks were generated and evaluated. Unlike other domains where this network topology has been used, in ultrasound there is low visual variation between classes. This work shows the potential for CNNs to be applied to this low-variation domain of medical imaging for disease discrimination.

The calculation of Cardiothoracic Ratio (CTR) in digital chest radiographs would be useful for cardiac anomaly assessment and heart enlargement related disease indication. The purpose of this study was to develop and evaluate a fully automated scheme for calculation of CTR in digital chest radiographs. Our automated method consisted of three steps, i.e., lung region localization, lung segmentation, and CTR calculation. We manually annotated the lung boundary with 84 points in 100 digital chest radiographs, and calculated an average lung model for the subsequent work. Firstly, in order to localize the lung region, generalized Hough transform was employed to identify the upper, lower, and outer boundaries of lung by use of Sobel gradient information. The average lung model was aligned to the localized lung region to obtain the initial lung outline. Secondly, we separately applied dynamic programming method to detect the upper, lower, outer and inner boundaries of lungs, and then linked the four boundaries to segment the lungs. Based on the identified outer boundaries of left lung and right lung, we corrected the center and the declination of the original radiography. Finally, CTR was calculated as a ratio of the transverse diameter of the heart to the internal diameter of the chest, based on the segmented lungs. The preliminary results on 106 digital chest radiographs showed that the proposed method could obtain accurate segmentation of lung based on subjective observation, and achieved sensitivity of 88.9% (40 of 45 abnormalities), and specificity of 100% (i.e. 61 of 61 normal) for the identification of heart enlargements.

Barretts Esophagus (BE) is a precancerous condition that affects the esophagus tube and has the risk of developing esophageal adenocarcinoma. BE is the process of developing metaplastic intestinal epithelium and replacing the normal cells in the esophageal area. The detection of BE is considered difficult due to its appearance and properties. The diagnosis is usually done through both endoscopy and biopsy. Recently, Computer Aided Diagnosis systems have been developed to support physicians opinion when facing difficulty in detection/classification in different types of diseases. In this paper, an automatic classification of Barretts Esophagus condition is introduced. The presented method enhances the internal features of a Confocal Laser Endomicroscopy (CLE) image by utilizing a proposed enhancement filter. This filter depends on fractional differentiation and integration that improve the features in the discrete wavelet transform of an image. Later on, various features are extracted from each enhanced image on different levels for the multi-classification process. Our approach is validated on a dataset that consists of a group of 32 patients with 262 images with different histology grades. The experimental results demonstrated the efficiency of the proposed technique. Our method helps clinicians for more accurate classification. This potentially helps to reduce the need for biopsies needed for diagnosis, facilitate the regular monitoring of treatment/development of the patients case and can help train doctors with the new endoscopy technology. The accurate automatic classification is particularly important for the Intestinal Metaplasia (IM) type, which could turn into deadly cancerous. Hence, this work contributes to automatic classification that facilitates early intervention/treatment and decreasing biopsy samples needed.

Computer-aided detection (CAD) of flat polyps, in contrast to other polyp types, is challenging due to their lack of projections from the colonic surface and limited geometrical features that can be extracted from such polyps. In this paper, we present a new approach for CAD of flat polyps via colon wall thickness mapping, texture feature extraction and analysis. First, we integrated our previous work of detecting flat polyp candidates via colon wall thickness mapping into this study for automated detection of initial polyp candidates (IPCs). The colon wall segmentation is established on a coupled level-set method after the lumen is electronically cleansed by a sophisticated statistical algorithm, which considers the partial volume effect to preserve the mucosa layer details. The IPC detection was performed based on the wall thickness local pattern. From each IPC volume, we extracted the 14 Haralick texture features and 16 additional features that were previously demonstrated to improve polyp classification performance. Then, we adopted the Rpackage “randomForest” to classify the features for false positive (FP) reduction. We evaluated our method via 16 patient datasets. The proposed scheme achieved a high capacity in terms of the well-known area under the curve value of 0.930. The FPs was reduced to less than 3 FPs/per polyp. The experiment results demonstrate the feasibility of our method in achieving computer aided detection of flat polyps, therefore, improving the screening capability of computed tomography cololongraphy.

Computer-aided detection (CAD) of colonic polyps plays an important role in advancing computed tomographic colonography (CTC) toward a screening modality. Detection of flat polyps is very challenging because of their plaquelike morphology with limited geometric features for detection purpose. In this paper, we present a novel scheme to automatically detect initial polyp candidates (IPCs) of flat polyp in CTC images. First, tagged materials in CTC images were automatically removed via the partial volume (PV) based electronic colon cleansing (ECC) strategy. We then propose a dual level set competition model to segment the volumetric colon wall from CTC images after ECC. In this model, we developed a comprehensive cost function which takes consideration of the essential characteristics of colon wall such as colon mucosa and weak boundaries, to simulate the mutual interference relationships among those compositions of the colon wall. Furthermore, we introduced a CAD scheme based on the thickness mapping of the colon wall. By tracing the gradient direction of the potential field between inner and outer borders of the colon wall, we focus on the local thickness measures for the detection of IPCs. The proposed CAD approach was validated on patient CTC scans with flat polyps. Experimental results indicate that the present scheme is very promising towards detection of colonic flat polyp candidates via CTC.

Dual-energy computed tomography is used increasingly in CT colonography (CTC). The combination of computer-aided detection (CADe) and dual-energy CTC (DE-CTC) has high clinical value, because it can detect clinically significant colonic lesions automatically at higher accuracy than does conventional single-energy CTC. While CADe has demonstrated its ability to detect small polyps, its performance is highly dependent on several factors, including the quality of CTC images and electronic cleansing (EC) of the images. The presence of artifacts such as beam hardening and image noise in ultra-low-dose CTC can produce incorrectly cleansed colon images that severely degrade the detection performance of CTC for small polyps. Also, CADe methods are very dependent on the quality of input images and the information about different tissues in the colon. In this work, we developed a novel method to calculate EC images using spectral information from DE-CTC data. First, the ultra-low dose dual-energy projection data obtained from a CT scanner are decomposed into two materials, soft tissue and the orally administered fecal-tagging contrast agent, to detect the location and intensity of the contrast agent. Next, the images are iteratively reconstructed while gradually removing the presence of tagged materials from the images. Our preliminary qualitative results show that the method can cleanse the contrast agent and tagged materials correctly from DE-CTC images without affecting the appearance of surrounding tissue.

The retinal vasculature is the only part of the blood circulation system that can be observed non-invasively using fundus cameras. Changes in the dynamic properties of retinal blood vessels are associated with many systemic and vascular diseases, such as hypertension, coronary heart disease and diabetes. The assessment of the characteristics of the retinal vascular network provides important information for an early diagnosis and prognosis of many systemic and vascular diseases. The manual analysis of the retinal vessels and measurement of quantitative biomarkers in large-scale screening programs is a tedious task, time-consuming and costly. This paper describes a reliable, automated, and efficient retinal health information and notification system (acronym RHINO) which can extract a wealth of geometric biomarkers in large volumes of fundus images. The fully automated software presented in this paper includes vessel enhancement and segmentation, artery/vein classification, optic disc, fovea, and vessel junction detection, and bifurcation/crossing discrimination. Pipelining these tools allows the assessment of several quantitative vascular biomarkers: width, curvature, bifurcation geometry features and fractal dimension. The brain-inspired algorithms outperform most of the state-of-the-art techniques. Moreover, several annotation tools are implemented in RHINO for the manual labeling of arteries and veins, marking optic disc and fovea, and delineating vessel centerlines. The validation phase is ongoing and the software is currently being used for the analysis of retinal images from the Maastricht study (the Netherlands) which includes over 10,000 subjects (healthy and diabetic) with a broad spectrum of clinical measurements

Early detection of glaucoma is important to slow down progression of the disease and to prevent total vision loss. We have been studying an automated scheme for detection of a retinal nerve fiber layer defect (NFLD), which is one of the earliest signs of glaucoma on retinal fundus images. In our previous study, we proposed a multi-step detection scheme which consists of Gabor filtering, clustering and adaptive thresholding. The problems of the previous method were that the number of false positives (FPs) was still large and that the method included too many rules. In attempt to solve these problems, we investigated the end-to-end learning system without pre-specified features. A deep convolutional neural network (DCNN) with deconvolutional layers was trained to detect NFLD regions. In this preliminary investigation, we investigated effective ways of preparing the input images and compared the detection results. The optimal result was then compared with the result obtained by the previous method. DCNN training was carried out using original images of abnormal cases, original images of both normal and abnormal cases, ellipse-based polar transformed images, and transformed half images. The result showed that use of both normal and abnormal cases increased the sensitivity as well as the number of FPs. Although NFLDs are visualized with the highest contrast in green plane, the use of color images provided higher sensitivity than the use of green image only. The free response receiver operating characteristic curve using the transformed color images, which was the best among seven different sets studied, was comparable to that of the previous method. Use of DCNN has a potential to improve the generalizability of automated detection method of NFLDs and may be useful in assisting glaucoma diagnosis on retinal fundus images.

Quantitative biomarkers for assessing the presence, severity, and progression of age-related macular degeneration (AMD) would benefit research, diagnosis, and treatment. This paper explores development of quantitative biomarkers derived from OCT imagery of the retina. OCT images for approximately 75 patients with Wet AMD, Dry AMD, and no AMD (healthy eyes) were analyzed to identify image features indicative of the patients’ conditions. OCT image features provide a statistical characterization of the retina. Healthy eyes exhibit a layered structure, whereas chaotic patterns indicate the deterioration associated with AMD. Our approach uses wavelet and Frangi filtering, combined with statistical features that do not rely on image segmentation, to assess patient conditions. Classification analysis indicates clear separability of Wet AMD from other conditions, including Dry AMD and healthy retinas. The probability of correct classification of was 95.7%, as determined from cross validation. Similar classification analysis predicts the response of Wet AMD patients to treatment, as measured by the Best Corrected Visual Acuity (BCVA). A statistical model predicts BCVA from the imagery features with R² = 0.846. Initial analysis of OCT imagery indicates that imagery-derived features can provide useful biomarkers for characterization and quantification of AMD: Accurate assessment of Wet AMD compared to other conditions; image-based prediction of outcome for Wet AMD treatment; and features derived from the OCT imagery accurately predict BCVA; unlike many methods in the literature, our techniques do not rely on segmentation of the OCT image. Next steps include larger scale testing and validation.

Identification of optic nerve head (ONH) is necessary in retinal image analysis to locate anatomical components such as fovea and retinal vessels in fundus images. In this study, we first worked on two different methods for preprocessing of images after that our main method was proposed for ONH detection in color fundus images. In the first preprocessing method, we did color space conversion, illumination equalization, and contrast enhancement and separately in the second method we applied top-hat transformation to an image. In the next step, Radon transform is applied to each of these two preprocessed fundus image to find candidates for the location of the ONH. Then, the accurate location was found using the minimum mean square error estimation. The accuracy of this method was approved by the results. Our method detected ONH correctly in 110 out of 120 images in our local database and 38 out of 40 color images in the DRIVE database by using Illumination equalization and contrast enhancement preprocessing. Moreover, by use of top-hat transformation our approach correctly detected the ONHs in 106 out of 120 images in the local database and 36 out of 40 images in the DRIVE set. In addition, Sensitivity and specificity of pixel base analysis of this algorithm seems to be acceptable in comparison with other methods.

In most cases, the pituitary tumor compresses optic chiasma and causes optic nerves atrophy, which will reflect in retina. In this paper, an Adaboost classification based method is first proposed to screen pituitary tumor from retinal spectral- domain optical coherence tomography (SD-OCT) image. The method includes four parts: pre-processing, feature extraction and selection, training and testing. First, in the pre-processing step, the retinal OCT image is segmented into 10 layers and the first 5 layers are extracted as our volume of interest (VOI). Second, 19 textural and spatial features are extracted from the VOI. Principal component analysis (PCA) is utilized to select the primary features. Third, in the training step, an Adaboost based classifier is trained using the above features. Finally, in the testing phase, the trained model is utilized to screen pituitary tumor. The proposed method was evaluated on 40 retinal OCT images from 30 patients and 30 OCT images from 15 normal subjects. The accuracy rate for the diseased retina was (85.00±16.58)% and the rate for normal retina was (76.68±21.34)%. Totally average accuracy of the Adaboost classifier was (81.43± 9.15)%. The preliminary results demonstrated the feasibility of the proposed method.

Tooth pulp atrophy occurs with increasing age. An age estimation procedure using dental cone beam computed tomography (CBCT) imaging was developed. Clinical dental CBCT images of 60 patients (aged from 20 to 80 years) were evaluated. The ratio of the cross-sectional area of the pulp cavity to the cross-sectional area of the tooth (pulp cavity ratio) was calculated. The pulp cavity ratio in the labio-lingual plane of the mandibular anterior teeth and the mesio-distal plane of the maxillary anterior teeth was strongly correlated with the patients’ age. The pulp cavity ratio of anterior teeth may be a useful parameter for estimating age.

This manuscript presents an analytical treatment on the feasibility of multi-scale Gabor filter bank response for non-invasive oral cancer pre-screening and detection in the long infrared spectrum. Incapability of present healthcare technology to detect oral cancer in budding stage manifests in high mortality rate. The paper contributes a step towards automation in non-invasive computer-aided oral cancer detection using an amalgamation of image processing and machine intelligence paradigms. Previous works have shown the discriminative difference of facial temperature distribution between a normal subject and a patient. The proposed work, for the first time, exploits this difference further by representing the facial Region of Interest(ROI) using multiscale rotation invariant Gabor filter bank responses followed by classification using Radial Basis Function(RBF) kernelized Support Vector Machine(SVM). The proposed study reveals an initial increase in classification accuracy with incrementing image scales followed by degradation of performance; an indication that addition of more and more finer scales tend to embed noisy information instead of discriminative texture patterns. Moreover, the performance is consistently better for filter responses from profile faces compared to frontal faces.This is primarily attributed to the ineptness of Gabor kernels to analyze low spatial frequency components over a small facial surface area. On our dataset comprising of 81 malignant, 59 pre-cancerous, and 63 normal subjects, we achieve state-of-the-art accuracy of 85.16% for normal v/s precancerous and 84.72% for normal v/s malignant classification. This sets a benchmark for further investigation of multiscale feature extraction paradigms in IR spectrum for oral cancer detection.

In large disasters, dental record plays an important role in forensic identification. However, filing dental charts for corpses is not an easy task for general dentists. Moreover, it is laborious and time-consuming work in cases of large scale disasters. We have been investigating a tooth labeling method on dental cone-beam CT images for the purpose of automatic filing of dental charts. In our method, individual tooth in CT images are detected and classified into seven tooth types using deep convolutional neural network. We employed the fully convolutional network using AlexNet architecture for detecting each tooth and applied our previous method using regular AlexNet for classifying the detected teeth into 7 tooth types. From 52 CT volumes obtained by two imaging systems, five images each were randomly selected as test data, and the remaining 42 cases were used as training data. The result showed the tooth detection accuracy of 77.4% with the average false detection of 5.8 per image. The result indicates the potential utility of the proposed method for automatic recording of dental information.

This paper proposes morphological descriptors representing the degree of skull deformity for craniosynostosis in head CT images and a hierarchical classifier model distinguishing among normal and different types of craniosynostosis. First, to compare deformity surface model with mean normal surface model, mean normal surface models are generated for each age range and the mean normal surface model is deformed to the deformity surface model via multi-level threestage registration. Second, four shape features including local distance and area ratio indices are extracted in each five cranial bone. Finally, hierarchical SVM classifier is proposed to distinguish between the normal and deformity. As a result, the proposed method showed improved classification results compared to traditional cranial index. Our method can be used for the early diagnosis, surgical planning and postsurgical assessment of craniosynostosis as well as quantitative analysis of skull deformity.

Abdominal obesity is strongly associated with a number of diseases and accurately assessment of subtypes of adipose tissue volume plays a significant role in predicting disease risk, diagnosis and prognosis. The objective of this study is to develop and evaluate a new computer-aided detection (CAD) scheme based on deep learning models to automatically segment subcutaneous fat areas (SFA) and visceral (VFA) fat areas depicting on CT images. A dataset involving CT images from 40 patients were retrospectively collected and equally divided into two independent groups (i.e. training and testing group). The new CAD scheme consisted of two sequential convolutional neural networks (CNNs) namely, Selection-CNN and Segmentation-CNN. Selection-CNN was trained using 2,240 CT slices to automatically select CT slices belonging to abdomen areas and SegmentationCNN was trained using 84,000 fat-pixel patches to classify fat-pixels as belonging to SFA or VFA. Then, data from the testing group was used to evaluate the performance of the optimized CAD scheme. Comparing to manually labelled results, the classification accuracy of CT slices selection generated by Selection-CNN yielded 95.8%, while the accuracy of fat pixel segmentation using Segmentation-CNN yielded 96.8%. Therefore, this study demonstrated the feasibility of using deep learning based CAD scheme to recognize human abdominal section from CT scans and segment SFA and VFA from CT slices with high agreement compared with subjective segmentation results.

Considering the weak edges in pancreas segmentation, this paper proposes a new solution which integrates more features of CT images by combining SLIC superpixels and interactive region merging. In the proposed method, Mahalanobis distance is first utilized in SLIC method to generate better superpixel images. By extracting five texture features and one gray feature, the similarity measure between two superpixels becomes more reliable in interactive region merging. Furthermore, object edge blocks are accurately addressed by re-segmentation merging process. Applying the proposed method to four cases of abdominal CT images, we segment pancreatic tissues to verify the feasibility and effectiveness. The experimental results show that the proposed method can make segmentation accuracy increase to 92% on average. This study will boost the application process of pancreas segmentation for computer-aided diagnosis system.

The development of a random forests machine learning technique is presented for fully-automated neck, chest, abdomen, and pelvis tissue segmentation of CT images using Trainable WEKA (Waikato Environment for Knowledge Analysis) Segmentation (TWS) plugin of FIJI (ImageJ, NIH). The use of a single classifier model to segment six tissue classes (lung, fat, muscle, solid organ, blood/contrast agent, bone) in the CT images is studied. An automated unbiased scheme to sample pixels from the training images and generate a balanced training dataset over the seven classes is also developed. Two independent training datasets are generated from a pool of 4 adult (>55 kg) and 3 pediatric patients (<;=55 kg) with 7 manually contoured slices for each patient. Classifier training investigated 28 image filters comprising a total of 272 features. Highly correlated and insignificant features are eliminated using Correlated Feature Subset (CFS) selection with Best First Search (BFS) algorithms in WEKA. The 2 training models (from the 2 training datasets) had 74 and 71 input training features, respectively. The study also investigated the effect of varying the number of trees (25, 50, 100, and 200) in the random forest algorithm. The performance of the 2 classifier models are evaluated on inter-patient intra-slice, intrapatient inter-slice and inter-patient inter-slice test datasets. The Dice similarity coefficients (DSC) and confusion matrices are used to understand the performance of the classifiers across the tissue segments. The effect of number of features in the training input on the performance of the classifiers for tissue classes with less than optimal DSC values is also studied. The average DSC values for the two training models on the inter-patient intra-slice test data are: 0.98, 0.89, 0.87, 0.79, 0.68, and 0.84, for lung, fat, muscle, solid organ, blood/contrast agent, and bone, respectively. The study demonstrated that a robust segmentation accuracy for lung, muscle and fat tissue classes. For solid-organ, blood/contrast and bone, the performance of the segmentation pipeline improved significantly by using the advanced capabilities of WEKA. However, further improvements are needed to reduce the noise in the segmentation.

This pilot study performs texture analysis on multiple magnetic resonance (MR) images of common renal masses for differentiation of renal cell carcinoma (RCC). Bounding boxes are drawn around each mass on one axial slice in T1 delayed sequence to use for feature extraction and classification. All sequences (T1 delayed, venous, arterial, pre-contrast phases, T2, and T2 fat saturated sequences) are co-registered and texture features are extracted from each sequence simultaneously. Random forest is used to construct models to classify lesions on 96 normal regions, 87 clear cell RCCs, 8 papillary RCCs, and 21 renal oncocytomas; ground truths are verified through pathology reports.

The highest performance is seen in random forest model when data from all sequences are used in conjunction, achieving an overall classification accuracy of 83.7%. When using data from one single sequence, the overall accuracies achieved for T1 delayed, venous, arterial, and pre-contrast phase, T2, and T2 fat saturated were 79.1%, 70.5%, 56.2%, 61.0%, 60.0%, and 44.8%, respectively. This demonstrates promising results of utilizing intensity information from multiple MR sequences for accurate classification of renal masses.

Nonalcoholic fatty liver disease (NAFLD) is prevalent and has a worldwide distribution now. Although ultrasound imaging technology has been deemed as the common method to diagnose fatty liver, it is not able to detect NAFLD in its early stage and limited by the diagnostic instruments and some other factors. B-scan image feature extraction of fatty liver can assist doctor to analyze the patient’s situation and enhance the efficiency and accuracy of clinical diagnoses. However, some uncertain factors in ultrasonic diagnoses are often been ignored during feature extraction. In this study, the nonalcoholic fatty liver rabbit model was made and its liver ultrasound images were collected by setting different Thermal index of soft tissue (TIS) and mechanical index (MI). Then, texture features were calculated based on gray level co-occurrence matrix (GLCM) and the impacts of TIS and MI on these features were analyzed and discussed. Furthermore, the receiver operating characteristic (ROC) curve was used to evaluate whether each feature was effective or not when TIS and MI were given. The results showed that TIS and MI do affect the features extracted from the healthy liver, while the texture features of fatty liver are relatively stable. In addition, TIS set to 0.3 and MI equal to 0.9 might be a better choice when using a computer aided diagnosis (CAD) method for fatty liver recognition.

Deep learning is a trending promising method in medical image analysis area, but how to efficiently prepare the input image for the deep learning algorithms remains a challenge. In this paper, we introduced a novel artificial multichannel region of interest (ROI) generation procedure for convolutional neural networks (CNN). From LIDC database, we collected 54880 benign nodule samples and 59848 malignant nodule samples based on the radiologists’ annotations. The proposed CNN consists of three pairs of convolutional layers and two fully connected layers. For each original ROI, two new ROIs were generated: one contains the segmented nodule which highlighted the nodule shape, and the other one contains the gradient of the original ROI which highlighted the textures. By combining the three channel images into a pseudo color ROI, the CNN was trained and tested on the new multichannel ROIs (multichannel ROI II). For the comparison, we generated another type of multichannel image by replacing the gradient image channel with a ROI contains whitened background region (multichannel ROI I). With the 5-fold cross validation evaluation method, the CNN using multichannel ROI II achieved the ROI based area under the curve (AUC) of 0.8823±0.0177, compared to the AUC of 0.8484±0.0204 generated by the original ROI. By calculating the average of ROI scores from one nodule, the lesion based AUC using multichannel ROI was 0.8793±0.0210. By comparing the convolved features maps from CNN using different types of ROIs, it can be noted that multichannel ROI II contains more accurate nodule shapes and surrounding textures.

To automatic detect lung nodules from CT images, we designed a two stage computer aided detection (CAD) system. The first stage is graph cuts segmentation to identify and segment the nodule candidates, and the second stage is convolutional neural network for false positive reduction. The dataset contains 595 CT cases randomly selected from Lung Image Database Consortium and Image Database Resource Initiative (LIDC/IDRI) and the 305 pulmonary nodules achieved diagnosis consensus by all four experienced radiologists were our detection targets. Consider each slice as an individual sample, 2844 nodules were included in our database. The graph cuts segmentation was conducted in a two-dimension manner, 2733 lung nodule ROIs are successfully identified and segmented. With a false positive reduction by a seven-layer convolutional neural network, 2535 nodules remain detected while the false positive dropped to 31.6%. The average F-measure of segmented lung nodule tissue is 0.8501.

A method for the detection of juxta-pleural lung nodules with radius ≤ 5mm in chest computed tomography images is proposed. The lung volume is segmented using region-growing and refined with morphological operations and active contours to include juxta-pleural nodules. Nodule candidates are searched slice-wise inside the lung volume segmentation. Solid nodules are detected by selecting an appropriate threshold inside a representative sliding window. Sub-solid and non-solid nodules are enhanced with a multiscale Laplacian-of-Gaussian filtering prior to their detection. Obvious non-nodule candidates, namely small blood vessels, are discarded using fixed rules. Then, a support vector machine with radial basis function is trained with the remaining candidates to further reduce the number of false positives (FPs). The final system sensitivity is 57.4% with 4 FPs/scan.The performance is similar or better than state-of-the-art methods, especially when considering the high number and small radius of the studied juxta-pleural nodules.

In this study, we address the problem of lesion classification in radiographic scans. We adapt image kernel functions to be applicable for high-resolution, grayscale images to improve the classification accuracy of a support vector machine. We take existing kernel functions inspired by the histogram of oriented gradients, and derive an approximation that can be evaluated in linear time of the image size instead of the original quadratic complexity, enabling highresolution input. Moreover, we propose a new variant inspired by the matched filter, to better utilize intensity space. The new kernels are improved to be scale-invariant and combined with a Gaussian kernel built from handcrafted image features. We introduce a simple multiple kernel learning framework that is robust when one of the kernels, in the current case the image feature kernel, dominates the others. The combined kernel is input to a support vector classifier. We tested our method on lesion classification both in chest radiographs and digital tomosynthesis scans. The radiographs originated from a database including 364 patients with lung nodules and 150 healthy cases. The digital tomosynthesis scans were obtained by simulation using 91 CT scans from the LIDC-IDRI database as input. The new kernels showed good separation capability: ROC AuC was in [0.827, 0.853] for the radiograph database and 0.763 for the tomosynthesis scans. Adding the new kernels to the image-feature-based classifier significantly improved accuracy: AuC increased from 0.958 to 0.967 and from 0.788 to 0.801 for the two applications.

The chest digital tomosynthesis(CDT) is recently developed medical device that has several advantage for diagnosing lung disease. For example, CDT provides depth information with relatively low radiation dose compared to computed tomography (CT). However, a major problem with CDT is the image artifacts associated with data incompleteness resulting from limited angle data acquisition in CDT geometry. For this reason, the sensitivity of lung disease was not clear compared to CT. In this study, to improve sensitivity of lung disease detection in CDT, we developed computer aided diagnosis (CAD) systems based on machine learning. For design CAD systems, we used 100 cases of lung nodules cropped images and 100 cases of normal lesion cropped images acquired by lung man phantoms and proto type CDT. We used machine learning techniques based on support vector machine and Gabor filter. The Gabor filter was used for extracting characteristics of lung nodules and we compared performance of feature extraction of Gabor filter with various scale and orientation parameters. We used 3, 4, 5 scales and 4, 6, 8 orientations. After extracting features, support vector machine (SVM) was used for classifying feature of lesions. The linear, polynomial and Gaussian kernels of SVM were compared to decide the best SVM conditions for CDT reconstruction images. The results of CAD system with machine learning showed the capability of automatically lung lesion detection. Furthermore detection performance was the best when Gabor filter with 5 scale and 8 orientation and SVM with Gaussian kernel were used. In conclusion, our suggested CAD system showed improving sensitivity of lung lesion detection in CDT and decide Gabor filter and SVM conditions to achieve higher detection performance of our developed CAD system for CDT.

Among the detected small nodules sized from 3 to 30mm in CT images, a significant portion is undetermined in terms of malignancy which needs biopsy or other follow-up means, resulting in excessive risk and cost. Therefore, predicting the malignancy of the nodules becomes a clinically desirable task. Based on the previous study of texture features extracted from gray-tone spatial-dependence matrices, this study aims to find more efficient texture features or image texture markers in discriminating the nodule malignancy. Two new image texture markers (median and variance) are proposed to classify the small nodules into different malignant levels, thus the risk prediction could be performed through image analysis. These two new image texture markers can minimize the effect of outliers in the feature series, thus can reduce the noise influence to the feature classification. Total 1,353 nodule samples selected from the Lung Image Database Consortium were used to evaluate the efficiency of the proposed new features. All the classification results are shown in the ROC curves and tabulated by the AUC values. The classification outcomes from (1) the most likely and likely benign nodules vs. the most likely and likely malignant nodules, (2) the most likely vs. likely benign nodules, and (3) the most likely vs. likely malignant nodules, are 0.9125±0.0096, 0.9239±0.0147, and 0.8888±0.0197, respectively, in terms of the largest AUC values. From the experimental outcomes on different malignant levels, the two new image texture markers from nodule volumetric CT image data have shown encouraging performance for the risk prediction.

Accurately predict the risk of disease progression and benefit of tyrosine kinase inhibitors (TKIs) therapy for stage IV non-small cell lung cancer (NSCLC) patients with activing epidermal growth factor receptor (EGFR) mutations by current staging methods are challenge. We postulated that integrating a classifier consisted of multiple computed tomography (CT) phenotypic features, and other clinicopathological risk factors into a single model could improve risk stratification and prediction of progression-free survival (PFS) of EGFR TKIs for these patients. Patients confirmed as stage IV EGFR-mutant NSCLC received EGFR TKIs with no resection; pretreatment contrast enhanced CT performed at approximately 2 weeks before the treatment was enrolled. A six-CT-phenotypic-feature-based classifier constructed by the LASSO Cox regression model, and three clinicopathological factors: pathologic N category, performance status (PS) score, and intrapulmonary metastasis status were used to construct a nomogram in a training set of 115 patients. The prognostic and predictive accuracy of this nomogram was then subjected to an external independent validation of 107 patients. PFS between the training and independent validation set is no statistical difference by Mann-Whitney U test (P = 0.2670). PFS of the patients could be predicted with good consistency compared with the actual survival. C-index of the proposed individualized nomogram in the training set (0·707, 95%CI: 0·643, 0·771) and the independent validation set (0·715, 95%CI: 0·650, 0·780) showed the potential of clinical prognosis to predict PFS of stage IV EGFR-mutant NSCLC from EGFR TKIs. The individualized nomogram might facilitate patient counselling and individualise management of patients with this disease.

Airway segmentation is an important step in analyzing chest CT volumes for computerized lung cancer detection, emphysema diagnosis, asthma diagnosis, and pre- and intra-operative bronchoscope navigation. However, obtaining an integrated 3-D airway tree structure from a CT volume is a quite challenging task. This paper presents a novel airway segmentation method based on intensity structure analysis and bronchi shape structure analysis in volume of interest (VOI). This method segments the bronchial regions by applying the cavity enhancement filter (CEF) to trace the bronchial tree structure from the trachea. It uses the CEF in each VOI to segment each branch and to predict the positions of VOIs which envelope the bronchial regions in next level. At the same time, a leakage detection is performed to avoid the leakage by analysing the pixel information and the shape information of airway candidate regions extracted in the VOI. Bronchial regions are finally obtained by unifying the extracted airway regions. The experiments results showed that the proposed method can extract most of the bronchial region in each VOI and led good results of the airway segmentation.

In this paper, retrieval accuracy of different types of pulmonary nodules is studied. The trainee radiologists could enrich their knowledge using the visual information of the retrieved nodules. In the proposed retrieval system, the pulmonary nodules are segmented using a semi-automated technique. Several 3D features are explored to improve the performance of the proposed retrieval system. A set of relevant shape and texture features is determined for efficient representation of the nodules in the feature space. The proposed CBIR system is validated on a data set of 542 nodules of Lung Image Database Consortium and Image Database Resource Initiative (LIDC/IDRI). The nodules with composite rank of malignancy “1”, “2” are considered as benign and “4”, “5” are considered as malignant. Considering top five retrieved images, the precision of the proposed retrieval system are 84.76%, 80.75%, and 80.34% for well-circumscribed, juxta-pleural, and juxtavascular nodules, respectively.

In this paper, we investigate the effect of the error criteria used during a training phase of the artificial neural network (ANN) on the accuracy of the classifier for classification of lung tissues affected with Interstitial Lung Diseases (ILD). Mean square error (MSE) and the cross-entropy (CE) criteria are chosen being most popular choice in state-of-the-art implementations. The classification experiment performed on the six interstitial lung disease (ILD) patterns viz. Consolidation, Emphysema, Ground Glass Opacity, Micronodules, Fibrosis and Healthy from MedGIFT database. The texture features from an arbitrary region of interest (AROI) are extracted using Gabor filter. Two different neural networks are trained with the scaled conjugate gradient back propagation algorithm with MSE and CE error criteria function respectively for weight updation. Performance is evaluated in terms of average accuracy of these classifiers using 4 fold cross-validation. Each network is trained for five times for each fold with randomly initialized weight vectors and accuracies are computed. Significant improvement in classification accuracy is observed when ANN is trained by using CE (67.27%) as error function compared to MSE (63.60%). Moreover, standard deviation of the classification accuracy for the network trained with CE (6.69) error criteria is found less as compared to network trained with MSE (10.32) criteria.

Volume calculation from the Computed Tomography (CT) lung lesions data is a significant parameter for clinical diagnosis. The volume is widely used to assess the severity of the lung nodules and track its progression, however, the accuracy and efficiency of previous studies are not well achieved for clinical uses. It remains to be a challenging task due to its tight attachment to the lung wall, inhomogeneous background noises and large variations in sizes and shape.

In this paper, we employ Halton low-discrepancy sequences to calculate the volume of the lung lesions. The proposed method directly compute the volume without the procedure of three-dimension (3D) model reconstruction and surface triangulation, which significantly improves the efficiency and reduces the complexity. The main steps of the proposed method are: (1) generate a certain number of random points in each slice using Halton low-discrepancy sequences and calculate the lesion area of each slice through the proportion; (2) obtain the volume by integrating the areas in the sagittal direction. In order to evaluate our proposed method, the experiments were conducted on the sufficient data sets with different size of lung lesions. With the uniform distribution of random points, our proposed method achieves more accurate results compared with other methods, which demonstrates the robustness and accuracy for the volume calculation of CT lung lesions. In addition, our proposed method is easy to follow and can be extensively applied to other applications, e.g., volume calculation of liver tumor, atrial wall aneurysm, etc.

Positron emission tomography (PET) imaging has been widely explored for treatment outcome prediction. Radiomicsdriven methods provide a new insight to quantitatively explore underlying information from PET images. However, it is still a challenging problem to automatically extract clinically meaningful features for prognosis. In this work, we develop a PET-guided distant failure predictive model for early stage non-small cell lung cancer (NSCLC) patients after stereotactic ablative radiotherapy (SABR) by using sparse representation. The proposed method does not need precalculated features and can learn intrinsically distinctive features contributing to classification of patients with distant failure. The proposed framework includes two main parts: 1) intra-tumor heterogeneity description; and 2) dictionary pair learning based sparse representation. Tumor heterogeneity is initially captured through anisotropic kernel and represented as a set of concatenated vectors, which forms the sample gallery. Then, given a test tumor image, its identity (i.e., distant failure or not) is classified by applying the dictionary pair learning based sparse representation. We evaluate the proposed approach on 48 NSCLC patients treated by SABR at our institute. Experimental results show that the proposed approach can achieve an area under the characteristic curve (AUC) of 0.70 with a sensitivity of 69.87% and a specificity of 69.51% using a five-fold cross validation.

Radiological imaging and image interpretation for clinical decision making are mostly specific to each body region such as head & neck, thorax, abdomen, pelvis, and extremities. For automating image analysis and consistency of results, standardizing definitions of body regions and the various anatomic objects, tissue regions, and zones in them becomes essential. Assuming that a standardized definition of body regions is available, a fundamental early step needed in automated image and object analytics is to automatically trim the given image stack into image volumes exactly satisfying the body region definition. This paper presents a solution to this problem based on the concept of virtual landmarks and evaluates it on whole-body positron emission tomography/computed tomography (PET/CT) scans. The method first selects a (set of) reference object(s), segments it (them) roughly, and identifies virtual landmarks for the object(s). The geometric relationship between these landmarks and the boundary locations of body regions in the craniocaudal direction is then learned through a neural network regressor, and the locations are predicted. Based on low-dose unenhanced CT images of 180 near whole-body PET/CT scans (which includes 34 whole-body PET/CT scans), the mean localization error for the boundaries of superior of thorax (TS) and inferior of thorax (TI), expressed as number of slices (slice spacing ≈ 4mm)), and using either the skeleton or the pleural spaces as reference objects, is found to be 3,2 (using skeleton) and 3, 5 (using pleural spaces) respectively, or in mm 13, 10 mm (using skeleton) and 10.5, 20 mm (using pleural spaces), respectively. Improvements of this performance via optimal selection of objects and virtual landmarks and other object analytics applications are currently being pursued.

and the skeleton and pleural spaces used as a reference objects

Purpose: To evaluate a new approach for predicting nodule volume measurement bias and variability when scanning with a specific CT scanner and acquisition protocol. Methods: A GE LightSpeed VCT scanner was used to scan 3 new rolls of 3M 3/4 x 1000 Inch Scotch Magic tape with a routine chest protocol (120 kVp, 100 mA, 0.4 s rotation, .98 pitch, STANDARD kernel) at three different slice thicknesses and spacings. Each tape scan was independently analyzed by fully automated image quality assessment software, producing fundamental image quality characteristics and simulated lung nodule volume measurements for a range of sphere diameters. The same VCT scanner and protocol was then used to obtain 10 repeat CT scans of an anthropomorphic chest phantom containing multiple Teflon spheres embedded in foam (diameters = 4.76mm, 6.25mm, and 7.94mm). The observed volume of the spheres in the 30 (3 reconstructions per scan) repeat scans was provided by independently developed nodule measurement software. Results: The predicted vs observed mean volume (mm3 ) and CV for 3 slice thicknesses and sphere sizes was obtained. For 0.625mm slice thickness scans the predicted vs observed values were (44.3,0.91)-vs-(48.2,1.17), (110.4,0.51)-vs-(124.1,0.47), and (219.9,0.29)-vs-(250.1,0.34), for 4.76mm, 6.25mm, and 7.94mm spheres respectively. For 1.25mm slice thickness the corresponding values were (42.1,0.98)-vs-(47.6,1.35), (106.9,0.56)-vs-(123.1,0.61), and (214.8,0.32)-vs-(248.8,0.41). For 2.5mm slice thickness the corresponding values were (23.9,9.53)-vs-(36.8,12.50), (77.6,3.84)-vs-(110.5,3.20), and (173.0,1.57)-vs-(233.9,1.32). Conclusion: Volume measurement bias and variability for lung nodules based on nodule size and acquisition protocol can potentially be predicted using a new method that utilizes fundamental image characteristics and simulation.

Creating classifiers for computer-aided diagnosis in the absence of ground truth is a challenging problem. Using experts’ opinions as reference truth is difficult because the variability in the experts’ interpretations introduces uncertainty in the labeled diagnostic data. This uncertainty translates into noise, which can significantly affect the performance of any classifier on test data. To address this problem, we propose a new label set weighting approach to combine the experts’ interpretations and their variability, as well as a selective iterative classification (SIC) approach that is based on conformal prediction. Using the NIH/NCI Lung Image Database Consortium (LIDC) dataset in which four radiologists interpreted the lung nodule characteristics, including the degree of malignancy, we illustrate the benefits of the proposed approach. Our results show that the proposed 2-label-weighted approach significantly outperforms the accuracy of the original 5- label and 2-label-unweighted classification approaches by 39.9% and 7.6%, respectively. We also found that the weighted 2-label models produce higher skewness values by 1.05 and 0.61 for non-SIC and SIC respectively on root mean square error (RMSE) distributions. When each approach was combined with selective iterative classification, this further improved the accuracy of classification for the 2-weighted-label by 7.5% over the original, and improved the skewness of the 5-label and 2-unweighted-label by 0.22 and 0.44 respectively.

The presence of a vertebral compression fracture is highly indicative of osteoporosis and represents the single most robust predictor for development of a second osteoporotic fracture in the spine or elsewhere. Less than one third of vertebral compression fractures are diagnosed clinically. We present an automated method for detecting spine compression fractures in Computed Tomography (CT) scans. The algorithm is composed of three processes. First, the spinal column is segmented and sagittal patches are extracted. The patches are then binary classified using a Convolutional Neural Network (CNN). Finally a Recurrent Neural Network (RNN) is utilized to predict whether a vertebral fracture is present in the series of patches.

Contrast Enhanced Ultrasound (CEUS) is a sensitive imaging technique to assess tissue vascularity, that can be useful in the quantification of different perfusion patterns. This can particularly important in the early detection and differentiation of different types of arthritis. A Gamma-variate can accurately quantify synovial perfusion and it is flexible enough to describe many heterogeneous patterns. However, in some cases the heterogeneity of the kinetics can be such that even the Gamma model does not properly describe the curve, especially in presence of recirculation or of an additional slowflow component.

In this work we apply to CEUS data both the Gamma-variate and the single compartment recirculation model (SCR) which takes explicitly into account an additional component of slow flow. The models are solved within a Bayesian framework.

We also employed the perfusion estimates obtained with SCR to train a support vector machine classifier to distinguish different types of arthritis. When dividing the patients into two groups (rheumatoid arthritis and polyarticular RA-like psoriatic arthritis vs. other arthritis types), the slow component amplitude was significantly different across groups: mean values of a₁ and its variability were statistically higher in RA and RA-like patients (131% increase in mean, p = 0.035 and 73% increase in standard deviation, p = 0.049 respectively). The SVM classifier achieved a balanced accuracy of 89%, with a sensitivity of 100% and a specificity of 78%.

Amyotrophic lateral sclerosis (ALS) causes functional disorders such as difficulty in breathing and swallowing through the atrophy of voluntary muscles. ALS in its early stages is difficult to diagnose because of the difficulty in differentiating it from other muscular diseases. In addition, image inspection methods for aggressive diagnosis for ALS have not yet been established. The purpose of this study is to develop an automatic analysis system of the whole skeletal muscle to support the early differential diagnosis of ALS using whole-body CT images. In this study, the muscular atrophy parts including ALS patients are automatically identified by recognizing and segmenting whole skeletal muscle in the preliminary steps. First, the skeleton is identified by its gray value information. Second, the initial area of the body cavity is recognized by the deformation of the thoracic cavity based on the anatomical segmented skeleton. Third, the abdominal cavity boundary is recognized using ABM for precisely recognizing the body cavity. The body cavity is precisely recognized by non-rigid registration method based on the reference points of the abdominal cavity boundary. Fourth, the whole skeletal muscle is recognized by excluding the skeleton, the body cavity, and the subcutaneous fat. Additionally, the areas of muscular atrophy including ALS patients are automatically identified by comparison of the muscle mass. The experiments were carried out for ten cases with abnormality in the skeletal muscle. Global recognition and segmentation of the whole skeletal muscle were well realized in eight cases. Moreover, the areas of muscular atrophy including ALS patients were well identified in the lower limbs. As a result, this study indicated the basic technology to detect the muscle atrophy including ALS. In the future, it will be necessary to consider methods to differentiate other kinds of muscular atrophy as well as the clinical application of this detection method for early ALS detection and examine a large number of cases with stage and disease type.

Lumbar disc degeneration (LDD) is a commonly age-associated condition related to low back pain, while its consequences are responsible for over 90% of spine surgical procedures. In clinical practice, grading of LDD by inspecting MRI is a necessary step to make a suitable treatment plan. This step purely relies on physicians manual inspection so that it brings the unbearable tediousness and inefficiency. An automated method for grading of LDD is highly desirable. However, the technical implementation faces a big challenge from class ambiguity, which is typical in medical image classification problems with a large number of classes. This typical challenge is derived from the complexity and diversity of medical images, which lead to a serious class overlapping and brings a great challenge in discriminating different classes. To solve this problem, we proposed an automated grading approach, which is based on supervised distance metric learning to classify the input discs into four class labels (0: normal, 1: slight, 2: marked, 3: severe). By learning distance metrics from labeled instances, an optimal distance metric is modeled and with two attractive advantages: (1) keeps images from the same classes close, and (2) keeps images from different classes far apart. The experiments, performed in 93 subjects, demonstrated the superiority of our method with accuracy 0.9226, sensitivity 0.9655, specificity 0.9083, F-score 0.8615. With our approach, physicians will be free from the tediousness and patients will be provided an effective treatment.

The segmentation and labeling of the individual bones serve as the first step to the fully automated measurement of skeletal characteristics and the detection of abnormalities such as skeletal deformities, osteoporosis, and vertebral fractures. Moreover, the identified landmarks on the segmented bone structures can potentially provide relatively reliable location reference to other non-rigid human organs, such as breast, heart and lung, thereby facilitating the corresponding image analysis and registration. A fully automated anatomy-directed framework for the segmentation and labeling of the individual bone structures from low-dose chest CT is presented in this paper. The proposed system consists of four main stages: First, both clavicles are segmented and labeled by fitting a piecewise cylindrical envelope. Second, the sternum is segmented under the spatial constraints provided by the segmented clavicles. Third, all ribs are segmented and labeled based on 3D region growing within the volume of interest defined with reference to the spinal canal centerline and lungs. Fourth, the individual thoracic vertebrae are segmented and labeled by image intensity based analysis in the spatial region constrained by the previously segmented bone structures. The system performance was validated with 1270 lowdose chest CT scans through visual evaluation. Satisfactory performance was obtained respectively in 97.1% cases for the clavicle segmentation and labeling, in 97.3% cases for the sternum segmentation, in 97.2% cases for the rib segmentation, in 94.2% cases for the rib labeling, in 92.4% cases for vertebra segmentation and in 89.9% cases for the vertebra labeling.

Osteoarthritis is the most common rheumatic disease in the world. Knee pain is the most disabling symptom in the disease, the prediction of pain is one of the targets in preventive medicine, this can be applied to new therapies or treatments. Using the magnetic resonance imaging and the grading scales, a multivariate model based on genetic algorithms is presented. Using a predictive model can be useful to associate minor structure changes in the joint with the future knee pain. Results suggest that multivariate models can be predictive with future knee chronic pain. All models; T0, T1 and T2, were statistically significant, all p values were < 0.05 and all AUC > 0.60.

Accurate lumbar spine measurement in CT images provides an essential way for quantitative spinal diseases analysis such as spondylolisthesis and scoliosis. In today’s clinical workflow, the measurements are manually performed by radiologists and surgeons, which is time consuming and irreproducible. Therefore, automatic and accurate lumbar spine measurement algorithm becomes highly desirable. In this study, we propose a method to automatically calculate five different lumbar spine measurements in CT images. There are three main stages of the proposed method: First, a learning based spine labeling method, which integrates both the image appearance and spine geometry information, is used to detect lumbar and sacrum vertebrae in CT images. Then, a multiatlases based image segmentation method is used to segment each lumbar vertebra and the sacrum based on the detection result. Finally, measurements are derived from the segmentation result of each vertebra. Our method has been evaluated on 138 spinal CT scans to automatically calculate five widely used clinical spine measurements. Experimental results show that our method can achieve more than 90% success rates across all the measurements. Our method also significantly improves the measurement efficiency compared to manual measurements. Besides benefiting the routine clinical diagnosis of spinal diseases, our method also enables the large scale data analytics for scientific and clinical researches.

Lower back pain and pathologies related to it are one of the most common results for a referral to a neurosurgical clinic in the developed and the developing world. Quantitative evaluation of these pathologies is a challenge. Image based measurements of angles/vertebral heights and disks could provide a potential quantitative biomarker for tracking and measuring these pathologies. Detection of vertebral bodies is a key element and is the focus of the current work. From the variety of medical imaging techniques, MRI and CT scans have been typically used for developing image segmentation methods. However, CT scans are known to give a large dose of x-rays, increasing cancer risk [8]. MRI can be substituted for CTs when the risk is high [8] but are difficult to obtain in smaller facilities due to cost and lack of expertise in the field [2]. X-rays provide another option with its ability to control the x-ray dosage, especially for young people, and its accessibility for smaller facilities. Hence, the ability to create quantitative biomarkers from x-ray data is especially valuable. Here, we develop a multiscale template matching, inspired by [9], to detect centers of vertebral bodies from x-ray data. The immediate application of such detection lies in developing quantitative biomarkers and in querying similar images in a database. Previously, shape similarity classification methods have been used to address this problem, but these are challenging to use in the presence of variation due to gross pathology and even subtle effects [1].

Lower back pain is one of the most prevalent disorders in the developed/developing world. However, its etiology is poorly understood and treatment is often determined subjectively. In order to quantitatively study the emergence and evolution of back pain, it is necessary to develop consistently measurable markers for pathology. Imaging based measures offer one solution to this problem. The development of imaging based on quantitative biomarkers for the lower back necessitates automated techniques to acquire this data. While the problem of segmenting lumbar vertebrae has been addressed repeatedly in literature, the associated problem of computing relevant biomarkers on the basis of the segmentation has not been addressed thoroughly. In this paper, we propose a Random-Forest based approach that learns to segment vertebral bodies in CT images followed by a biomarker evaluation framework that extracts vertebral heights and widths from the segmentations obtained. Our dataset consists of 15 CT sagittal scans obtained from General Electric Healthcare. Our main approach is divided into three parts: the first stage is image pre-processing which is used to correct for variations in illumination across all the images followed by preparing the foreground and background objects from images; the next stage is Machine Learning using Random-Forests, which distinguishes the interest-point vectors between foreground or background; and the last step is image post-processing, which is crucial to refine the results of classifier. The Dice coefficient was used as a statistical validation metric to evaluate the performance of our segmentations with an average value of 0.725 for our dataset.

Predicting metastatic tumor response to chemotherapy at early stage is critically important for improving efficacy of clinical trials of testing new chemotherapy drugs. However, using current response evaluation criteria in solid tumors (RECIST) guidelines only yields a limited accuracy to predict tumor response. In order to address this clinical challenge, we applied Radiomics approach to develop a new quantitative image analysis scheme, aiming to accurately assess the tumor response to new chemotherapy treatment, for the advanced ovarian cancer patients. During the experiment, a retrospective dataset containing 57 patients was assembled, each of which has two sets of CT images: pre-therapy and 4-6 week follow up CT images. A Radiomics based image analysis scheme was then applied on these images, which is composed of three steps. First, the tumors depicted on the CT images were segmented by a hybrid tumor segmentation scheme. Then, a total of 115 features were computed from the segmented tumors, which can be grouped as 1) volume based features; 2) density based features; and 3) wavelet features. Finally, an optimal feature cluster was selected based on the single feature performance and an equal-weighed fusion rule was applied to generate the final predicting score. The results demonstrated that the single feature achieved an area under the receiver operating characteristic curve (AUC) of 0.838±0.053. This investigation demonstrates that the Radiomic approach may have the potential in the development of high accuracy predicting model for early stage prognostic assessment of ovarian cancer patients.

Prostatic adenocarcinoma is one of the most commonly occurring cancers among men in the world, and it also the most curable cancer when it is detected early. Multiparametric MRI (mpMRI) combines anatomic and functional prostate imaging techniques, which have been shown to produce high sensitivity and specificity in cancer localization, which is important in planning biopsies and focal therapies. However, in previous investigations, lesion localization was achieved mainly by manual segmentation, which is time-consuming and prone to observer variability. Here, we developed an algorithm based on locality alignment discriminant analysis (LADA) technique, which can be considered as a version of linear discriminant analysis (LDA) localized to patches in the feature space. Sensitivity, specificity and accuracy generated by the proposed algorithm in five prostates by LADA were 52.2%, 89.1% and 85.1% respectively, compared to 31.3%, 85.3% and 80.9% generated by LDA. The delineation accuracy attainable by this tool has a potential in increasing the cancer detection rate in biopsies and in minimizing collateral damage of surrounding tissues in focal therapies.

Voxel-based tissue outcome prediction in acute ischemic stroke patients is highly relevant for both clinical routine and research. Previous research has shown that features extracted from baseline multi-parametric MRI datasets have a high predictive value and can be used for the training of classifiers, which can generate tissue outcome predictions for both intravenous and conservative treatments. However, with the recent advent and popularization of intra-arterial thrombectomy treatment, novel research specifically addressing the utility of predictive classi- fiers for thrombectomy intervention is necessary for a holistic understanding of current stroke treatment options. The aim of this work was to develop three clinically viable tissue outcome prediction models using approximate nearest-neighbor, generalized linear model, and random decision forest approaches and to evaluate the accuracy of predicting tissue outcome after intra-arterial treatment. Therefore, the three machine learning models were trained, evaluated, and compared using datasets of 42 acute ischemic stroke patients treated with intra-arterial thrombectomy. Classifier training utilized eight voxel-based features extracted from baseline MRI datasets and five global features. Evaluation of classifier-based predictions was performed via comparison to the known tissue outcome, which was determined in follow-up imaging, using the Dice coefficient and leave-on-patient-out cross validation. The random decision forest prediction model led to the best tissue outcome predictions with a mean Dice coefficient of 0.37. The approximate nearest-neighbor and generalized linear model performed equally suboptimally with average Dice coefficients of 0.28 and 0.27 respectively, suggesting that both non-linearity and machine learning are desirable properties of a classifier well-suited to the intra-arterial tissue outcome prediction problem.

In abdominal disease diagnosis and various abdominal surgeries planning, segmentation of abdominal blood vessel (ABVs) is a very imperative task. Automatic segmentation enables fast and accurate processing of ABVs. We proposed a fully automatic approach for segmenting ABVs through contrast enhanced CT images by a hybrid of 3D region growing and 4D curvature analysis. The proposed method comprises three stages. First, candidates of bone, kidneys, ABVs and heart are segmented by an auto-adapted threshold. Second, bone is auto-segmented and classified into spine, ribs and pelvis. Third, ABVs are automatically segmented in two sub-steps: (1) kidneys and abdominal part of the heart are segmented, (2) ABVs are segmented by a hybrid approach that integrates a 3D region growing and 4D curvature analysis. Results are compared with two conventional methods. Results show that the proposed method is very promising in segmenting and classifying bone, segmenting whole ABVs and may have potential utility in clinical use.

Studies in intravascular optical coherence tomography (IV-OCT) have demonstrated the importance of coronary bifurcation regions in intravascular medical imaging analysis, as plaques are more likely to accumulate in this region leading to coronary disease. A typical IV-OCT pullback acquires hundreds of frames, thus developing an automated tool to classify the OCT frames as bifurcation or non-bifurcation can be an important step to speed up OCT pullbacks analysis and assist automated methods for atherosclerotic plaque quantification. In this work, we evaluate the performance of two state-of-the-art classifiers, SVM and Neural Networks in the bifurcation classification task. The study included IV-OCT frames from 9 patients. In order to improve classification performance, we trained and tested the SVM with different parameters by means of a grid search and different stop criteria were applied to the Neural Network classifier: mean square error, early stop and regularization. Different sets of features were tested, using feature selection techniques: PCA, LDA and scalar feature selection with correlation. Training and test were performed in sets with a maximum of 1460 OCT frames. We quantified our results in terms of false positive rate, true positive rate, accuracy, specificity, precision, false alarm, f-measure and area under ROC curve. Neural networks obtained the best classification accuracy, 98.83%, overcoming the results found in literature. Our methods appear to offer a robust and reliable automated classification of OCT frames that might assist physicians indicating potential frames to analyze. Methods for improving neural networks generalization have increased the classification performance.

Cone Beam Computed Tomography (CBCT), a kind of face and neck exams can be opportunity to identify, as an incidental finding, calcifications of the carotid artery (CACA). Given the similarity of the CACA with calcification found in several x-ray exams, this work suggests that a similar technique designed to detect breast calcifications in mammography images could be applied to detect such calcifications in CBCT. The method used a 3D version of the calcification detection technique [1], based on a signal enhancement using a convolution with a 3D Laplacian of Gaussian (LoG) function followed by removing the high contrast bone structure from the image. Initial promising results show a 71% sensitivity with 0.48 false positive per exam.

Deep learning has shown a great potential in computer aided diagnosis. However, in many applications, large dataset is not available. This makes the training of a sophisticated deep learning neural network (DNN) difficult. In this study, we demonstrated that with transfer learning, we can quickly retrain start-of-the-art DNN models with limited data provided by the prostateX challenge. The training data consists of 330 lesions, only 78 were clinical significant. Efforts were made to balance the data during training. We used ImageNet pre-trained inceptionV3 and Vgg-16 model and obtained AUC of 0.81 and 0.83 respectively on the prostateX test data, good for a 4th place finish. We noticed that models trained for different prostate zone has different sensitivity. Applying scaling factors before merging the result improves the AUC for the final result.

An overview of the 2017 PROSTATEx Challenge.

Overview of the Digital Mammography DREAM Challenge.

A participant's experience of the Digital Mammography DREAM Challenge.

A participant's experience in the Digital Mammography DREAM Challenge.