We define image quality by how accurately an observer, human or otherwise, can perform a given task, such as determining to which class an image belongs. For detection tasks, the Bayesian ideal observer is the best observer, in that it sets an upper bound for observer performance, summarized by the area under the receiver operating characteristic curve. However, the use of this observer is frequently infeasible because of unknown image statistics, whose estimation is computationally costly. As a result, a channelized ideal observer (CIO) was investigated to reduce the dimensionality of the data, yet approximate the performance of the ideal observer. Previously investigated channels include Laguerre Gauss (LG) channels and channels via the singular value decomposition of the given linear system (SVD). Though both types are highly efficient for the ideal observer, they nevertheless have the weakness that they may not be as efficient for general detection tasks involving complex/realistic images; the former is particular to the signal and background shape, and the latter is particular to the system operator. In this work, we attempt to develop channels that can be applied to a system with any signal and background type and without knowledge of any characteristics of the system. The method used is a partial least squares algorithm (PLS), in which channels are chosen to maximize the squared covariance between images and their classes. Preliminary results show that the CIO with PLS channels outperforms one with either the LG or SVD channels and very closely approximates ideal-observer performance.

Date Published: 12 March 2009
PDF: 8 pages
Proc. SPIE 7263, Medical Imaging 2009: Image Perception, Observer Performance, and Technology Assessment, 72630Q (12 March 2009); doi: 10.1117/12.813550

Published in SPIE Proceedings Vol. 7263:
Medical Imaging 2009: Image Perception, Observer Performance, and Technology Assessment
Berkman Sahiner; David J. Manning, Editor(s)