Integrated region and edge segmentation using information divergence measures

Abstract

Listen

Translate article

Accurate models of human anatomy are obligatory for modern cancer radiotherapy. Maps of individual patient anatomies are usually drawn manually from CT images. Manual contouring is expensive and time-consuming because of the complexity of the anatomy, the low contrast of soft tissues in CT, and blurred image detail due to respiratory motion. We have developed automated contouring methods based on relative entropy and more general divergence measures from information theory and statistics that produce average minimum error inference like the traditional maximum likelihood (ML) and maximum a posteriori (MAP) classifiers. Unlike the ML/MAP classifiers that are frequently implemented assuming Gaussian models for the data, the information theoretic divergences require no data model. We have concentrated on the Jensen-Renyi divergence (JRD) by which multiple contours can be obtained simultaneously with the optimization of a single objective function. Region segmentation is accomplished by maximizing the divergence of pixel feature distributions inside and outside a flexible, closed, parametric curve. Recently we have integrated multivariate region segmentation with edge detection, also done by maximizing the JRD over sets of region-interior and region-edge pixels in edge-enhanced versions of the image. Further, region and edge detection are combined with prior shape constraints in which the combined JRD objective function is penalized if the flexible curve parameters deviate too far from those of a prior known shape. Though the performance of the JRD program is a complex function of pixel feature number and kind, edges, and shape priors, we demonstrate accurate contours computed from image data distributions and estimates of prior shape.