Detection of multiple sclerosis lesions in MRI's with neural networks

The objective of this paper is to state the effectiveness of a two-stage learning classification system in the automatic detection of small lesions from Magnetic Resonance Images (MRIs) of a patient affected by multiple sclerosis. The first classification stage consists of an unsupervised neural network module for data clustering. The second classification stage consists of a supervised learning module employing a plurality vote mechanism to relate each unsupervised cluster to the supervised output class having the largest number of representatives inside the cluster. In this paper two different neural network algorithms, i.e. the Enhanced Linde-Buzo-Gray (ELBG) algorithm and the well-known Self-Organizing Map (SOM), have been employed as the clustering module in the first stage of the system, respectively. The results obtained with the two different clustering algorithms have been qualitatively and quantitatively compared in a set of classification experiments. In these experiments, ELBG is equivalent to SOM in terms of classification accuracy and superior to SOM with respect to the visual quality of the output map and robustness to changes in the order and composition of the data presentation sequence. The results confirm the usefulness of the neural classification system in the automatic the detection of small lesions.