An adaptive sparse Bayesian model combined with joint information-based label fusion for brain tumor segmentation in MRI

This paper focuses on the automatic segmentation of brain tumor in MR image. The first focus is an algorithm that applies the sparse Bayesian decision theorem combined with the Markov random field and Gibbs random field for the segmentation of brain tumor in T1-w, T1-c, T2-w and FLAIR image, respectively. The second focus is the joint label fusion algorithm based on intensity information and spatial information. We employ the spatial information of inner voxels in reliable regions to help the segmentation of boundary voxels in non-reliable regions in label fusion process. The proposed approach is evaluated by 20 training cases obtained from Medical Image Computing and Computer Aided Intervention Society (MICCAI) 2017 Brain Tumor Segmentation Challenge. A comparison of the proposed approach and other conventional approaches is presented in terms of Dice, Sensitivity and Hausdorff_95. Furthermore, the total tumor volume is calculated and compared with expert delineation. It can be seen that the proposed approach is able to acquire a larger mean value of Dice and Sensitivity than the other conventional approaches. This novel approach has an added value for the clinical evaluation of tumor patients.