Multicriteria fuzzy clustering for brain image segmentation

One of the most challenging task in image analysis is to identify correctly tissues where boundaries are generally not clear. Fuzzy clustering is supposed to be the most appropriate to model this situation in applications such as tissue classification, tumor detection. While, image segmentation using fuzzy clustering technique classifies correctly pixels of an image with a great extent of accuracy [1], recent works have shown that fuzzy clustering techniques considers a single objective may not provide a good result since no single validity measure works well on different kinds of data sets. Moreover, a wrong choice of a validity measure leads to poor results [2]. In this paper, we introduce a multiobjective fuzzy clustering approach producing a set of Pareto solutions among which the best solution, based on I-index validation measure, is chosen to be the final clustering solution. First, a spatial information is considered to deal more effectively with the noise and intensity inhomogeneities introduced in imaging process. Second, we propose to use a variable string length encoding technique to automatically identify the number of clusters, given that it does not require a prior knowledge about number of clusters present in a data set. Therefore, an initializing method based on a center approximation approach is proposed to accelerate the clustering process and make results more robust. Applied to normal and multiple sclerosis lesion magnetic resonance image brain images, our method shows better performance than competing algorithms.