Multi-Modal Brain Tumor Segmentation via Multi-Category Interaction and Graph Co-Reasoning

Numerous multi-modal tumor segmentation techniques have been developed to identify affected regions in brain images, thereby enhancing the intelligence of diagnosis. Nevertheless, the morphological and semantic knowledge of brain tumors in multi-modal images remains incompletely explored, and spatial relations between multiple categories are still under-researched. The limited exploration of multi-modal images leads to an irrational distribution of tumor areas in the predictive results. To address this issue, this work proposes a Graph Co-reasoning Network (GCoN), in which the morphological and semantic information between regions and contours interact to guide the graph reasoning, facilitating the recognition of tumor areas. The GCoN also incorporates a Multi-category Interaction Module (MIM), which models the dependency between multiple categories, thus enabling more accurate subregion localization of brain tumors. In addition, we derive a unified graph re-projection, which re-projects the results of intra- and inter-Graph Attention Network (GAT) into the prediction space through a unified assignment matrix. Experimental results demonstrate that our method outperforms the current state-of-the-art methods, achieving 87.54%, 87.76% and 90.02% average Dice scores on the BraTS2019, BraTS2020, and BraTS2021 datasets, respectively.