Deep Cascaded Attention Network for Multi-task Brain Tumor Segmentation

Abstract

Multi-modal magnetic resonance images (MRIs) play an important role in the diagnosis and treatment of brain tumors. Due to heterogeneous diversities, it’s of great challenge to segment gliomas into hierarchical regions. Decomposing the multi-class segmentation task into sequential subtasks with cascaded models has proved its effectiveness, but leads to model redundancy, training complexity, and task isolation. In this paper, we propose a simple yet efficient 3D deep cascaded attention network (DCAN) for brain tumor segmentation. Specifically, we settle multi-tasks into corresponding branches with a shared feature extractor to reduce model complexity. Second, instead of explicitly extracting spatial evolutionary relationships of sub-regions using several consecutive models, a cascaded attention mechanism is introduced to implicitly involve potential subregions correlations as guidance. Moreover, we present a feature bridge module (FBM) to narrow feature fusion gaps. Thus, DCAN is able to capture the hierarchical correlations of overlapping regions and simultaneously tackle multi-tasks in a single model. The comprehensive experimental comparisons on the BRATS 2018 dataset show DCAN achieves top performance with dice scores of 81.71%, 91.18% and 86.19% for the enhancing tumor, whole tumor and tumor core, respectively.