Research Information System
IEEE Access, vol.14, pp.23996-24017, 2026 (SCI-Expanded, Scopus)
Medical image segmentation is a critical task for many clinical applications; however, the process of obtaining high-quality pixel-level annotations is both costly and time-consuming. Semi-supervised learning (SSL) has been proposed to alleviate the annotation cost of the model by utilizing large volumes of unlabeled data. However, existing SSL segmentation methods, such as pseudo-labeling and perturbation-based consistency regularization often suffer from low supervision quality because the pseudo-labels may be noisy and pixel-wise consistency fails to capture the complex anatomical context and boundaries of medical structures. To overcome these challenges, we propose a collaborative dual network training architecture named CDR-Net which utilizes contrastive regional-level knowledge distillation to enhance segmentation performance. Our method involves two parallel segmentation networks that work together by sharing regional-level knowledge in both intra- and inter-network contexts. Specifically, the proposed region-based contrastive distillation mechanism enhances cross-network knowledge transfer by aligning anatomically corresponding regions and penalizing inconsistent predictions, thereby providing more reliable regional supervision from unlabeled data. Unlike previous SSL segmentation methods that mainly concentrate on network perturbations or direct pixel-wise consistency constraints, our framework incorporates region-level contrastive distillation to enhance cross-network knowledge transfer, thereby boosting segmentation robustness and generalization. Comprehensive experiments conducted on three public datasets namely Left Atrium (LA) Segmentation Challenge Dataset, the NIH Pancreas Dataset and the Hippocampus Dataset show that our approach effectively leverages both labeled and unlabeled data, outperforming the state-of-the-art semi-supervised segmentation networks for complex anatomical structures. The code is avaliable at: https://github.com/ridvan25/CDR-Net.