SAResU-Net: Shuffle attention residual U-Net for brain tumor segmentation

Computer-aided segmentation technology is important for clinical treatment of brain tumors. In recent years, U-shaped networks have become mainstream for medical image segmentation, significantly improving the performance of brain tumor segmentation tasks. Since merits of the U -shaped architecture, we propose a new shuffle attention residual U-Net, i.e., SAResU-Net, for brain tumor segmentation application. SAResU-Net combines several shuffle attention (SA) blocks and residual modules with a basic 3D U-Net, where SA blocks are added to skip connection positions to capture the local spatial and channel information. In addition, a self-ensemble module is leveraged to further boost the model performance. Evaluation experimental results on the 2019 and 2020 Brain Tumor Segmentation (BraTS) datasets show that our SAResU-Net is superior to its baseline, especially on the tumor core segmentation task. Moreover, our model achieves DSC values of 79.17%, 90.02% and 82.00% for the enhancing tumor (ET), the whole tumor (WT), and tumor core(TC) on the BraTS 2020 validation dataset, respectively, while on the validation dataset of BraTS 2019, the values are 77.74%, 90.40% and 83.58%, respectively, proving its effectiveness in the application of brain tumor segmentation.