Neural inhabitants of MR and echo images segment cardiac structures

Abstract

Describes a new approach to the problem of the segmentation of cardiac structures in medical imaging. The approach is based on the idea of breeding and selecting artificial creatures who live in such images and are fed with the boundaries of the structures to be segmented. The authors' creatures, the Gnets, are simple individuals based on recurrent neural networks who can see, know their position in the environment, move inside it and eat. Their behavior is developed through a genetic algorithm which keeps a population of Gnets and mates the best individuals. Performance is evaluated on a set of test images of known segmentation. Preliminary results of this approach are reported.<>