Brain structures segmentation using optimum global and local weights on mixing active contours and neighboring constraints

Abstract

This paper presents a new method for segmenting multiple brain structures by using an optimized mixture of different Active Contour Models (ACMs). Prior constraints and structures' neighboring interaction are modelled for each structure. Prior information is also captured by a training process, in which structure's dependent local and global weights are calculated. The local weights regulate locally the combination of each term during the evolution, acting as an experienced balancer between image and prior information. The ideal proportion of relation between the mixture of different ACMs and the prior model is defined by the optimum global weights. As proof of concept, the method is applied on the very challenging task of segmenting hippocampus and amygdala structures.