Learning to Segment Using Machine-Learned Penalized Logistic Models

Abstract

Classical maximum-a-posteriori (MAP) segmentation uses generative models for images. However, creating tractable generative models can be difficult for complex images. Moreover, generative models require auxiliary parameters to be included in the maximization, which makes the maximization more complicated. This paper proposes an alternative to the MAP approach: using a penalized logistic model to directly model the segmentation posterior. This approach has two advantages: (1) It requires fewer auxiliary parameters, and (2) it provides a standard way of incorporating powerful machine-learning methods into segmentation so that complex image phenomenon can be learned easily from a training set. The technique is used to segment cardiac ultrasound images sequences which have substantial spatio-temporal contrast variation that is cumbersome to model. Experimental results show that the method gives accurate segmentations of the endocardium in spite of the contrast variation.