Harmonizing Generalization and Specialization: Uncertainty-Informed Collaborative Learning for Semi-Supervised...

Area of Science:

• Computer Vision
• Medical Imaging
• Machine Learning

Background:

• Vision foundation models offer strong generalization for medical image segmentation.
• These models struggle with specialized tasks due to limited annotations and rare pathologies, stemming from a mismatch between general priors and specific needs.

Purpose of the Study:

• To introduce Uncertainty-informed Collaborative Learning (UnCoL), a dual-teacher framework for semi-supervised medical image segmentation.
• UnCoL aims to harmonize generalization and specialization for improved performance on clinical tasks.

Main Methods:

• UnCoL employs a dual-teacher framework, distilling general knowledge from a frozen foundation model and task-specific knowledge from an adapting teacher.
• Pseudo-label learning is adaptively regulated by predictive uncertainty to stabilize learning in ambiguous regions and suppress unreliable supervision.

Main Results:

• UnCoL consistently outperforms existing semi-supervised methods across diverse 2D and 3D medical image segmentation benchmarks.
• The model achieves near fully supervised performance, demonstrating significantly reduced annotation requirements.

Conclusions:

• Uncertainty-informed Collaborative Learning effectively bridges the gap between general pretraining and specialized clinical requirements in medical image segmentation.
• UnCoL offers a promising solution for efficient and accurate medical image segmentation with limited labeled data.