Automatic segmentation of low resolution fetal cardiac data using snakes with shape priors

This paper presents a level set deformable model to segment all four chambers of the fetal heart simultaneously. We show its results in 2D on 53 images taken from only 8 datasets. Due to our lack of sufficient data we built only a mean template from the training data instead of a full active shape model. Using rigid registration the template was registered to unseen images and the snakes were guided by individual chamber priors as they evolved in unison to segment missing cardiac structures in the presence of high noise. Using a leave one out approach most of the segmentation errors are within 3 pixels of manually traced contours.