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

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Semi-supervised learning is crucial for medical image segmentation but can suffer from cognitive bias due to limited labeled data.
• This bias is amplified in complex medical images, leading to unstable predictions in ambiguous regions.
• Lack of supervision diversity further hinders self-correction and exploration of alternative predictions.

Purpose of the Study:

• To mitigate cognitive bias in semi-supervised medical image segmentation.
• To improve the accuracy and robustness of segmentation models trained with limited labeled data.

Main Methods:

• Proposed a framework combining Alternating Training and Pseudo Label Correction.
• Alternating Training uses original and augmented data to prevent biased reinforcement in ambiguous regions.
• Pseudo Label Correction refines unstable pseudo-labels using perturbation-driven consistency and cross-subnet constraints.

Main Results:

• Achieved superior performance on three public datasets with only 10% labeled data.
• Demonstrated significant improvements in Dice scores and reductions in 95HD compared to strong baselines.
• Effectively mitigated bias accumulation and boosted overall segmentation performance.

Conclusions:

• The framework successfully identifies and rectifies fragile predictions during training.
• Progressively refines unstable pseudo-labels, alleviating cognitive bias in semi-supervised medical image segmentation.
• Code is publicly available for reproducibility.