Texture based MRI segmentation with a two-stage hybrid neural classifier

We propose an automated method for extracting anatomical structures in magnetic resonance images (MRI) based on texture classification. It consists of two consecutive stages. The textures of an input MRI are first classified by a network spline neurons, organized within a hybrid master classifier/mixtures-of-experts architecture (stage I). The output map is then fed into a second neural network, which aims to better contrast the target structure and eliminate the mistakes of the first phase via local shape/texture analysis and a carefully designed learning process (stage II). Results are demonstrated on medical imagery with the segmentation of various brain structures.