CSSD: Cross-Supervision and Self-denoising for Hybrid-Supervised Hepatic Vessel Segmentation

Qiuting Hu, Li Lin, Pujin Cheng, Xiaoying Tang

Abstract: Hepatic vessel segmentation from Computer Tomography (CT) plays a crucial role in the diagnosis and treatment of various diseases. However, manually delineating hepatic vessels is a time-consuming, arduous task that demands expertise, rendering the procurement of substantial, high-fidelity annotated data from experts a challenge. Public available datasets often comprise unlabelled data or labels afflicted by noise. Recent research efforts have focused on utilizing these unlabelled data or noisy labels to extract additional information. However, these methods typically involve single-level supervision, lacking the joint use of unlabelled data and noisy labels. Therefore, developing a model training paradigm that can effectively combine multiple hierarchical supervision levels is desirable. To address this issue, we propose a novel framework that robustly learns segmentation from a small amount of relatively high-quality labelled data, a large amount of noisy labelled data and unlabelled data through Hybrid-Supervision Learning (HSL). Specifically, we employ two parallel segmentation networks to learn from unlabelled data using perturbation consistency of pseudo labels, and introduce the Cross-Sample Mutual Attention (CMA) module to transfer prior knowledge of relatively high-quality labelled data to unlabelled data, and employ a self-denoising approach to address potential issues with model generalization stemming from inaccuracies within noisy labels. Extensive experiments on public datasets (3DIRCADb, MSD3, MSD8) demonstrate the superiority of the proposed framework.