Two-Stage Segmentation Framework with Parallel Decoders for the Kidney and Kidney Tumor Segmentation

Kidney cancer is one of the most common cancers worldwide. Automated segmentation of kidneys, tumors, and cysts from CT images is an important pathway to assist doctors in diagnosis. However, the diverse morphologies of tumors and cysts pose challenges regarding their difficult identification and unpredictable behavior. This paper proposes a two-stage segmentation method, proceeding from coarse to fine. In the first stage, we obtain the kidney region of interest (ROI) based on nnU-Net as input for the second stage. In the second stage, we design a parallel encoder structure. It employs a dual-stream end-to-end training approach, simultaneously monitoring and segmenting boundary information and targets. In particular, a residual channel attention mechanism was incorporated with the boundary prediction branch, highlighting the most relevant feature channels. The method has been experimentally demonstrated to be significantly superior to the baseline nnU-Net. On the official test set, our Kidney + Masses Dice and Tumor Dice are 0.936 and 0.670, respectively, ranking 14th on the leaderboard.