Segmentation of glioblastoma multiforme from multimodal MR images based on superpixel

A superpixel-based segmentation method is proposed in this paper to improve the accuracy and robustness of GBM segmentation. In this approach, first, a local k-means clustering utilizing weighted distance is developed to over-segment the MR images into a series of superpixels that are not only homogeneous, compact but also match image boundaries. Then, a dynamic region merging algorithm based on sequential probability ratio test (SPRT) is performed to progressively integrate the neighboring superpixels. Finally, the GBM tissues are extracted using clustering algorithm. It is worth noting that the region merging algorithm used in this paper can preserve certain global properties, so that the results are neither over-merged nor under-merged. Experiments based on the images collected from 15 GBM patients were carried out to evaluate our proposed algorithm. Comparative results demonstrated that the proposed algorithm outperformed the FCM-based and the normalized cut (Ncut) algorithms.