GMM-PA: Gaussian Mixture Model-Based Prototype Alignment for Multi-source Domain Adaptation in Polyp Segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate polyp segmentation in colonoscopy is crucial for early colorectal cancer detection.
• Deep learning models face performance issues in cross-center colonoscopy data due to domain shift.

Purpose of the Study:

• To develop a robust multi-source domain adaptation network for cross-domain polyp segmentation.
• To enhance the performance of polyp segmentation models on unseen data from different centers.

Main Methods:

• Proposed a Gaussian Mixture Model-guided Prototype Alignment (GMM-PA) network.
• Implemented an image preprocessing module for domain discrepancy mitigation (LAB color transformation, Fourier transform spectral exchange).
• Utilized a hybrid CNN-Mamba architecture for feature extraction and employed adversarial learning for domain-invariant features.

Main Results:

• The GMM-PA method demonstrated superior performance in cross-domain polyp segmentation.
• Achieved an average Dice score of 0.8396 and mIoU of 0.8449 across three public datasets.
• Outperformed existing state-of-the-art approaches in cross-domain settings.

Conclusions:

• The proposed GMM-PA network effectively addresses domain shift in colonoscopy polyp segmentation.
• The method offers a robust solution for improving colorectal cancer detection through enhanced image analysis.
• The hybrid CNN-Mamba architecture and GMM-guided alignment contribute to superior cross-domain generalization.