A Dual Domain Collaborative Network for Polyp Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate polyp segmentation in colonoscopy is crucial for early colorectal cancer detection.
• Existing methods struggle with indistinct boundaries and scale variations, limiting segmentation accuracy.

Purpose of the Study:

• To introduce a novel dual domain collaborative network (DDCNet) for improved polyp segmentation.
• To address limitations in boundary feature optimization and cross-level representation alignment.

Main Methods:

• Developed a frequency context enhancement module (FCEM) for refining frequency domain features.
• Introduced a cross-level shift recalibrated fusion module (CSFM) for spatial domain feature alignment.
• Designed a hybrid loss function combining boundary, cross-entropy, and frequency consistency losses.

Main Results:

• DDCNet achieved state-of-the-art performance on three benchmark datasets (Kvasir SEG, CVC-ClinicDB, CVC-ColonDB).
• Achieved Dice coefficients of 0.9343, 0.9447, and 0.8155, surpassing existing methods by 1.0%-1.5%.
• Ablation studies confirmed the effectiveness of FCEM, CSFM, and the hybrid loss function.

Conclusions:

• DDCNet effectively improves polyp segmentation accuracy by enhancing boundary features and cross-level alignment.
• The proposed method offers a significant advancement for automated polyp detection in colonoscopy.
• The hybrid loss function and novel modules contribute to superior segmentation performance.