GSCA-Net: A Global Spatial Channel Attention Network for Kidney, Tumor and Cyst Segmentation

Automatic segmentation of the kidney, tumor, and cysts is crucial for the treatment of renal cancer. In this paper, we employed a 3D residual U-Net architecture as the pre-processing method to extract the region of interest (ROI) and segment the kidney. Then, we propose Global Spatial Channel Attention Network (GSCA-Net) with global spatial attention (GSA) and global channel attention (GCA) for the segmentation of tumors and cysts. Global spatial attention improves the global spatial representation ability, and global channel attention learns features between different channels. The GSCA module enhances the segmentation accuracy of tumors and cysts through the fusion of two parallel global attention modules. Furthermore, we employ a novel boundary loss function in GSCA-Net to improve the Surface Dice. On the official test set including cases 589–698, our approach achieves Dice coefficients of 0.933, 0.744, and 0.679 for the kidney, masses, and tumor, respectively.