Domain-Interactive Contrastive Learning and Prototype-Guided Self-Training for Cross-Domain Polyp Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate polyp segmentation in colonoscopy is crucial for colorectal cancer diagnosis and treatment.
• Deep learning models struggle with performance degradation on datasets from different imaging devices.
• Unsupervised Domain Adaptation (UDA) methods aim to bridge the domain gap using labeled source and unlabeled target data.

Purpose of the Study:

• To propose a novel framework, Domain-interactive Contrastive Learning and Prototype-guided Self-training (DCL-PS), for cross-domain polyp segmentation.
• To address limitations of existing UDA methods, including neglect of domain-wise representations and pseudo-label uncertainty.

Main Methods:

• Domain-interactive Contrastive Learning (DCL) with a domain-mixed prototype updating strategy to differentiate cross-domain class-wise features.
• Contrastive learning-based cross-consistency training (CL-CCT) to enhance encoder feature extraction.
• Prototype-guided Self-training (PS) with dynamic pixel weighting to improve pseudo-label quality.

Main Results:

• The proposed DCL-PS framework demonstrates superior performance in polyp segmentation on target domain datasets.
• The method effectively reduces the domain gap and improves model generalization.
• Experimental results validate the effectiveness of the DCL and PS strategies.

Conclusions:

• The DCL-PS framework offers a robust solution for cross-domain polyp segmentation in colonoscopy images.
• The proposed strategies enhance feature discrimination and pseudo-label reliability.
• This work contributes to more accurate and reliable polyp detection for colorectal cancer screening.