Refinement-based noise-conditioned learning for semi-supervised medical image segmentation

Semi-supervised medical image segmentation aims to reduce annotation costs by exploiting unlabeled data, yet its performance is often limited by unreliable pseudo-label supervision. In existing studies, V-Net and its variants are predominantly adopted as backbone architectures, and most advances focus on improving training strategies rather than exploring architectural diversity. We observe that directly applying classical segmentation models to semi-supervised settings, without task-specific configuration adjustment, often leads to overfitting when models are trained to predict segmentation masks from scratch. Interestingly, our empirical analysis reveals that reformulating the same segmentation models as refinement learners substantially improves performance under unreliable pseudo-label supervision.Based on this observation, we propose Refinement-based Noise-conditioned Learning (RNCL), a general framework for semi-supervised medical image segmentation. RNCL trains a refinement model to improve coarse predictions generated by a guide model, and further regulates the influence of guidance by injecting controlled Gaussian noise into the guide predictions during training. This noise-conditioned formulation exposes the refinement model to guidance signals with different degrees of corruption, encouraging adaptive integration of image information and guidance rather than trivial copying.In addition, RNCL incorporates a Feature Fusion with Adaptive Gating (FFAG) module to integrate image features, joint image-guidance features, and feature-level representations from the guide model. Extensive experiments on three public medical image segmentation benchmarks demonstrate that RNCL consistently outperforms state-of-the-art semi-supervised methods under low-label settings, while ablation studies confirm the effectiveness and robustness of each component. The code is accessible at https://github.com/woaierxu/RNCL.