A clustering-based method for brain tumor segmentation

Abstract

This work introduces a new brain tumor segmentation method based on a new criterion function for threshold selection. The method aims to accomplish some features that are desirable in medical practice such as simplicity, speed, accuracy, and independency of user inputs or parameters. Although the criterion function can be seen as an extension of Otsu's criterion, we did not assume low intra-variance for both classes, object and background. The new criterion is adapted to treat the tumor as the object to be segmented and a portion of healthy brain tissue as the background. In order to speed up the search for the thresholds, the segmentation method exploits domain knowledge by using some feature points in the histogram to bound a range of interest for searching. We evaluated the method on a set of 30 patient cases with their respective magnetic resonance images and compared the criterion function against other well-known criteria for threshold selection. The new criterion outperformed the other criteria in segmenting the complete region affected by the tumor for each patient case.