This study investigates the segmentation ability of unsupervised clustering of the image feature space. A self-organizing map, a feed-forward neural network and a k-nearest neighbor classifier were compared in labeling brain slices from magnetic resonance imaging. Qualitative and quantitative tests were carried out using brain images of a patient with an infarction. Five different tissue classes were partitioned: white matter, gray matter, cerebrospinal fluid, fluid in the infarct region and gray matter in the infarct region. The SOM based method performed best in all the cases that were investigated. Especially, the stability of the method concerning the influence of the training set was superior.
reprint (pdf, 373k)