Dual Graph Reasoning Unit for Brain Tumor Segmentation

With the rapid development of deep learning, many fully automatic segmentation models are developed to solve the challenge of brain tumor segmentation. However, few methods focus on the rich relational information and contextual dependencies in multimodal MR images. In this paper, we propose a novel approach, called Dual Graph Reasoning Unit (DGRUnit), for brain tumor segmentation. The proposed method consists of two parallel modules, a spatial reasoning module, and a channel reasoning module. The spatial reasoning module maps the original features to an embedding spatial node space and employs the graph convolutional network (GCN) to capture long-range relations between different regions in the spatial dimension. Similar to the spatial reasoning module, the channel reasoning module adopts the graph attention network (GAT) to model the rich contextual interdependencies between different channels with similar semantics representing. To demonstrate the effectiveness of our proposed method, we integrate both modules into a Nested U-net. Experimental results show that our approach achieve significant improvement on brain tumor segmentation task compared to several state-of-the-art methods.