Medical Image Segmentation by Partitioning Spatially Constrained Fuzzy Approximation Spaces

Abstract

Image segmentation is an important prerequisite step for any automatic clinical analysis technique. It assists in visualization of human tissues, as accurate delineation of medical images requires involvement of expert practitioners, which is also time consuming. In this background, the rough-fuzzy clustering algorithm provides an effective approach for image segmentation. It handles uncertainties arising due to overlapping classes and incompleteness in class definition by partitioning the fuzzy approximation spaces. However, the existing rough-fuzzy clustering algorithms do not consider the spatial distribution of the image. They depend only on the distribution of pixels to determine their class labels. In this regard, this article introduces a new algorithm, termed as spatially constrained rough-fuzzy $c$ -means (sRFCM) for medical image segmentation. The proposed sRFCM algorithm combines wisely the merits of rough-fuzzy clustering and local neighborhood information. In the proposed algorithm, the labels of local neighbors influence in the determination of the label of center pixel. The effect of local neighbors acts as a regularizer. Moreover, the proposed sRFCM algorithm partitions each cluster in possibilistic lower approximation or core region and probabilistic boundary region. The cluster centroid depends on the core and boundary regions, weight parameter, and neighborhood regularizer. A novel segmentation validity index, termed as neighborhood Silhouette, is proposed to find out the optimum values of regularizer and weight parameter, controlling the performance of the sRFCM. The efficacy of the proposed sRFCM algorithm, as well as several existing segmentation algorithms, is demonstrated on four brain MR volume databases and one HEp-2 cell image data.