Rethinking Copy-Paste for Consistency Learning in Medical Image Segmentation.

Area of Science:

• Medical image analysis
• Computer vision
• Machine learning

Background:

• Semi-supervised learning (SSL) with consistency learning shows promise for medical image segmentation.
• Existing copy-paste data augmentation methods can degrade synthetic label accuracy and over-perturb training data distributions, hindering generalization.

Purpose of the Study:

• To develop a robust weak-to-strong consistency learning framework for medical image segmentation.
• To address limitations of traditional copy-paste methods by enhancing synthetic label quality and controlling data perturbations.

Main Methods:

• Proposing a framework using cross-copy-pasting between labeled and unlabeled data to improve synthetic label reliability.
• Implementing uncertainty estimation and foreground region constraints to guide copy-pasting, ensuring beneficial perturbations.

Main Results:

• The framework significantly outperforms state-of-the-art models across six diverse medical image segmentation datasets.
• Achieved a 15.31% higher Dice score on the PROMISE12 prostate segmentation task using only 10% labeled data.

Conclusions:

• The proposed method effectively enhances medical image segmentation by refining the copy-paste technique within a consistency learning framework.
• The approach demonstrates superior generalization and performance, particularly in low-data regimes, offering a valuable tool for medical image analysis.