An artificial neural network modular system for the classification of anatomical objects in cerebral magnetic resonance images

An artificial modular system to get an anatomical objects classification is presented. The study analyzes cerebral images obtained by MR in normal subjects. The system consists of two modules based on neural architectures joined in sequence to perform an image segmentation followed by an anatomical objects classification. These two trials were performed with a Self Organizing Map and a Multilayer Perceptron trained with the Back Propagation learning rule respectively. The goal of the system is the automatic recognition of the anatomical structures in MR images of the cerebral section passing through the orbits and the visual pathways. To reach this purpose the Authors submitted the two networks to a training phase realized by an unsupervised process for the image segmentation and by a supervised process for regions labelling. This last step was based on topographic relations supplied by an expert neuroradiologist. The system was useful in discriminating 20 different classes of anatomical objects on the considered section. The results of the study are presented, and the data are examinated also in terms of efficiency and purity.