Mutual-Prototype Adaptation 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 for polyp segmentation struggle with domain shifts, degrading performance on new datasets.
• Manual annotation for new datasets is time-consuming and resource-intensive, necessitating domain adaptation techniques.

Purpose of the Study:

• To develop a robust method for automatic polyp segmentation that overcomes domain shifts in colonoscopy images.
• To leverage knowledge from labeled source domains to improve segmentation performance in unlabeled target domains.
• To enhance the generalizability of deep learning models for polyp segmentation across multi-center and multi-device data.

Main Methods:

• Proposed a mutual-prototype adaptation network featuring a mutual-prototype alignment (MPA) module for feature refinement.
• Introduced progressive self-training (PST) with an uncertainty-guided loss to generate accurate target domain prototypes.
• Incorporated a disentangled reconstruction (DR) module to maintain semantic consistency and provide complementary supervision.

Main Results:

• The proposed model effectively eliminates domain shifts in multi-center and multi-device colonoscopy images.
• Experimental results on CVC-DB, Kvasir-SEG, and ETIS-Larib datasets demonstrate superior polyp segmentation performance.
• The model significantly outperforms existing state-of-the-art methods in segmentation accuracy.

Conclusions:

• The mutual-prototype adaptation network offers a promising solution for domain-adaptive polyp segmentation.
• The integration of MPA, PST, and DR modules enhances model robustness and generalization.
• This approach holds potential for improving the clinical utility of automated polyp detection in colonoscopy.