Clique identification and propagation for multimodal brain tumor image segmentation

Brain tumors vary considerably in size, morphology, and location across patients, thus pose great challenge in automated brain tumor segmentation methods. Inspired by the concept of clique in graph theory, we present a clique-based method for multimodal brain tumor segmentation that considers a brain tumor image as a graph and automatically segment it into different sub-structures based on the clique homogeneity. Our proposed method has three steps, neighborhood construction, clique identification, and clique propagation. We constructed the neighborhood of each pixel based on its similarities to the surrounding pixels, and then extracted all cliques with a certain size k to evaluate the correlations among different pixels. The connections among all cliques were represented as a transition matrix, and a clique propagation method was developed to group the cliques into different regions. This method is also designed to accommodate multimodal features, as multimodal neuroimaging data is widely used in mapping the tumor-induced changes in the brain. To evaluate this method, we conduct the segmentation experiments on the publicly available Multimodal Brain Tumor Image Segmentation Benchmark (BRATS) dataset. The qualitative and quantitative results demonstrate that our proposed clique-based method achieved better performance compared to the conventional pixel-based methods.