Pseudo LABELIng Based 3DResUNet with DEep Supervision for the Kidney PArsing Segmentation Challenge

Abstract

Three-dimensional (3D) integrated renal structures (IRS) segmentation is an important task based on computed tomography angiography (CTA) images for laparoscopic partial nephrectomy (LPN). Fine renal artery segmentation based on a CTA image is an important step for kidney disease diagnosis and pre-operative planning. However, it could be challenging due to large inter-anatomy variation, thin structures, small volume ratio, and small labeled dataset of the fine renal artery. The Kidney PArsing Challenge (KiPA) 2022 organizers launched a challenge for automatic segmentation of kidneys, renal tumors, arteries, and veins. In this paper, we have proposed a two-stage solution for the automatic segmentation of four kidney-related structures on CTA images. In the first stage, the 3DResUNet with deep supervision has been used to generate pseudo labels. The pseudo labels generated using validation images from the first stage along with training samples are used in nnUNet to obtain the final segmentation. The proposed solution produced an optimal performance for Multi-Structure Segmentation for Renal Cancer Treatment. The code will be publicly available at: https://github.com/RespectKnowledge/Pseudo-labeling_Segmentation_KiPA22_challenege.