Dual uncertainty-guided multi-model pseudo-label learning for semi-supervised medical image segmentation.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Semi-supervised medical image segmentation is crucial for clinical applications.
• Pseudo-label learning, a common technique, suffers from unstable training and low-quality pseudo-labels due to sample variations and noise accumulation.
• Existing methods struggle with the inherent uncertainties in pseudo-label generation.

Purpose of the Study:

• To propose a novel framework, dual uncertainty-guided multi-model pseudo-label learning (DUMM), to address the limitations of traditional pseudo-labeling in semi-supervised medical image segmentation.
• To enhance the stability and accuracy of the segmentation training process.
• To generate high-quality pseudo-labels by leveraging dual uncertainty mechanisms.

Main Methods:

• Developed a sample selection module based on sample-level uncertainty (SUS) for a stable training process.
• Implemented a multi-model pseudo-label generation module based on pixel-level uncertainty (PUM) for high-quality pseudo-label generation.
• Validated the DUMM framework on the ACDC2017 and ISIC2018 medical image datasets.

Main Results:

• The DUMM framework achieved significant improvements in Dice scores, with increases of 6.5% on ACDC2017 and 4.0% on ISIC2018 compared to the baseline.
• Demonstrated superior performance over comparative semi-supervised segmentation methods.
• The proposed approach effectively mitigates issues related to pseudo-label quality and training instability.

Conclusions:

• The dual uncertainty-guided multi-model pseudo-label learning framework (DUMM) is a feasible and effective approach for semi-supervised medical image segmentation.
• The integration of sample-level and pixel-level uncertainty improves both training stability and segmentation accuracy.
• DUMM offers a promising direction for advancing semi-supervised learning in medical imaging.