Neural Classification of Multiple Sclerosis Lesions in MR Images

A connectionist architecture is presented that is able to perform focal lesions classification in MR images of brain tissues affected by multiple sclerosis disease. Images are first segmented using a fuzzy technique; then the images of each cluster are processed in order to classify and label non-pathologic tissues and to locate all possible candidates to be sclerosis lesions. Finally the neural classification step is performed together with an estimate of the position and the shape for each lesion. Classification is implemented using a multi-layer perceptron as a way to code the prior knowledge about clinical features of a sclerosis lesion and to obtain fast performance. The network has been trained using a back-propagation scheme. The outline of the whole architecture is presented and the experimental results are reported.