A Rule-based Fuzzy Segmentation System with Automatic Generation of Membership Functions for Pathological Brain Mr Images

In this paper, we present a rule-based fuzzy segmentation system that is capable of segmenting magnetic resonance (MR) images of diseased human brains into physiologically and pathologically meaningful regions for display and measurement. We have developed a novel technique to automatically generate the membership functions for the fuzzy sets in the antecedent of the IF-THEN fuzzy rules in our system. Furthermore, our system incorporated anatomical knowledge about brain structures and lesions. More speciically, our system used the distance between pixels and ventricle boundary as a fuzzy property of periventricular hy-perintensity. Using this fuzzy system, we have performed two diierent types of segmentation tasks: (1) segmentation of brain images without lesions into three classes (grey matter, white matter and cerebrospinal uid); and (2) segmentation of brain images with periventricular le-sions into four classes. Fourteen brain images were processed by our rule-based system as well as by the standard fuzzy c-means (FCM) algorithm used for performance comparison. The results , connrmed by the medical experts, showed that the rule-based fuzzy system signiicantly outperformed the standard FCM in the segmentation of the abnormal brain images while the methods performed almost equally for the normal brain images. 2