An automatic lesion segmentation method for fast spin echo magnetic resonance images using an ensemble of neural networks

Abstract

Multiple sclerosis (MS) is a chronic disease of the central nervous system which attacks the insulating myelin coating of nerve fibers in the brain and spinal cord, leaving scar tissue which can be seen on magnetic resonance imaging (MRI) scans. There is a well recognised need for a robust, objective, accurate and reproducible automatic method for identifying multiple sclerosis lesions on proton density (PD) and T/sub 2/-weighted MRI. Feed-forward neural networks (FFNN) are computational techniques inspired by the physiology of the brain and used in the approximation of general mappings from one finite dimensional space to another. They present a practical application of the theoretical resolution of Hilbert's 13th problem by Kolmogorov and Lorenz, and have been used with success in a variety of applications. We present a method for automatic MS lesion segmentation for fast spin echo (FSE) images (PD-weighted & T/sub 2/-weighted) based on an ensemble of feed-forward neural networks. The FFNN of the input layer of the ensemble are trained with different portions of example lesion and non-lesion data which have previously been hand-segmented by a clinician. The final output of the ensemble is determined by a gate FFNN which is trained to weigh the response of the input layer to unseen training data. The ensemble was trained with data from 14 MS patients and evaluated with data from another 6. The results are presented.