CoInNet: A Convolution-Involution Network With a Novel Statistical Attention for Automatic Polyp Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Gastrointestinal polyps are common and early diagnosis aids in reducing colorectal cancer risk.
• Automated polyp segmentation systems assist surgeons but face challenges due to polyp variability and unclear boundaries.
• Existing deep learning models (CNNs, Vision Transformers) have limitations in capturing diverse visual patterns and feature dependencies.

Purpose of the Study:

• To propose CoInNet, a novel polyp segmentation model.
• To leverage convolution and involution operations for enhanced feature extraction.
• To improve the accuracy of polyp segmentation, especially for small and easily overlooked polyps.

Main Methods:

• Developed CoInNet, a polyp segmentation model with a novel feature extraction mechanism.
• Integrated convolution and involution operations to capture diverse visual patterns.
• Introduced a statistical feature attention unit to learn relationships between feature maps.
• Incorporated an anomaly boundary approximation module with recursive feature fusion to refine segmentation.

Main Results:

• CoInNet precisely segments polyps, including tiny ones occupying only 0.01% of an image area.
• The model demonstrates superior performance in highlighting polyp regions and learning accurate boundaries.
• CoInNet outperformed thirteen state-of-the-art methods on five benchmark polyp segmentation datasets.

Conclusions:

• CoInNet offers a significant advancement in automated polyp segmentation.
• The model's ability to detect small polyps is crucial for clinical applications, particularly in analyzing large datasets like wireless capsule endoscopy videos.
• CoInNet shows great potential for improving early diagnosis and reducing the risk of colorectal cancer.