Semi-supervised Abdominal Multi-organ Segmentation via Contour Aware Dual-Task Consistency

Abdominal multi-organ segmentation has great significance for clinical works such as improving disease diagnosis, healing analysis, and treatment planning. However, abdominal organs suffer from complexities such as different shapes, sizes, and blurred boundaries, and these make annotation have to waste a lot of time and labor. Therefore, this paper proposes a Contour Aware Dual-task Consistent (CADTC) semi-supervised framework for abdominal multi-organ segmentation. This framework can learn diverse-level information from two distinct task branches, making it suitable for the intricate segmentation of abdominal organs, and Semi-Supervised Segmentation (SSS) is fitting for training with limited labeled data. The novelty of this framework is the proposed Contour Aware VNet (CA-VNet) for generating higher-quality prediction maps. The Contour Aware Module (CAM) integrates geometric shape information into the feature extraction process. The ablation experiments are conducted on CA-VNet and the overall model accuracy is found to be significantly improved, proving its effectiveness in fusing global detail information and geometric shape information. Experimental results on two datasets for abdominal multi-organ segmentation show that the proposed method outperforms other SOTA works.