SEMI-PLC: A framework for semi-supervised medical images segmentation with pseudo label correction

Semi-supervised learning has become a promising paradigm for medical image segmentation by leveraging both labeled and unlabeled data. However, neural networks trained with limited labeled samples often suffer from cognitive bias, where unstable predictions in uncertain regions reinforce bias over time, particularly in medical images with complex textures and ambiguous boundaries. This issue may be further exacerbated by the lack of supervision diversity, which limits the model's ability to self-correct and explore alternative predictions.

To mitigate cognitive bias in semi-supervised medical image segmentation, we propose a framework that jointly modeling inter-subnet prediction discrepancies and intra-subnet instability. First, we design an Alternating Training strategy on labeled data, where original and CutMix-augmented images are alternately fed into two subnets across training iterations. This alternation introduces supervision signal discrepancies that help prevent biased reinforcement in ambiguous regions, encouraging the model to explore alternative semantic interpretations. Second, we propose a Pseudo Label Correction mechanism for unlabeled data, which leverages perturbation-driven consistency to identify and suppress unstable pseudo labels. Specifically, CutMix serves as a perturbation-driven probe to reveal unstable predictions within subnets, which are progressively refined through cross-subnet consistency constraints. Combined with Alternating Training, this mechanism mitigates bias accumulation, and boosts the overall performance of semi-supervised medical image segmentation.

Experiments on three public dataset demonstrate that our method achieves superior performance with only 10% labeled data, improving Dice scores by 1.12, 2.87, and 0.32, and reducing 95HD by 1.59 mm, 3.25 mm, and 0.88 mm, respectively, compared with strong baselines.

The proposed framework identifies and rectifies fragile predictions during training, progressively refining unstable pseudo labels and alleviating cognitive bias in semi-supervised medical image segmentation. Code is available at: https://github.com/Shiyuan-H/SEMI-PLC.