Two-stage Cascaded Network With Deep Supervision And Residual Attention For Brain Tumor Segmentation

Accurate segmentation of brain tumor has important research value and significance for tumor growth evaluation and treatment. In this paper, we propose a novel two-stage cascaded network based on ResUNet, which is used to segment brain tumors and their substructures to solve the class imbalance problem in brain tumor data. Brain tumor segmentation includes the following two stages. In the first stage, we use a variant of U-Net with deep supervision to segment the whole tumor structure and construct a tumor mask for more tumor boundary information. In the second stage, we design a deep residual network which combines context extractor and attention mechanism to segment tumor substructure, aiming to focus on small target areas. Among them, the tumor mask of the first network is multiplied with the multimodal images as the input of the second network. The whole network is trained in an end-to-end fashion. The proposed method is evaluated on the BraTS 2018 training dataset and achieve competitive performance compared with the state-of-the-art approaches. The results of mean dice are 0.8813, 0.8056 and 0.7366 for the whole tumor, tumor core and enhancing tumor respectively.