Bilateral collaborative streams with multi-modal attention network for accurate polyp segmentation.

Area of Science:

• Medical Imaging and Artificial Intelligence
• Computational Pathology
• Gastroenterology

Background:

• Accurate segmentation of colorectal polyps in colonoscopy is crucial for early cancer detection and prevention.
• Existing segmentation methods face challenges with polyp diversity (size, morphology, texture) and computational efficiency for clinical use.

Purpose of the Study:

• To propose a novel dual-stream deep learning architecture, Bilateral Convolutional Multi-Attention Network (BiCoMA), for accurate colorectal polyp segmentation.
• To enhance segmentation performance by effectively integrating global contextual information and local spatial details.

Main Methods:

• Developed a dual-stream architecture (BiCoMA) combining convolutional neural networks (ConvNeXt V2 Large) and vision transformers (Pyramid Vision Transformer).
• Integrated Spatial Refinement (SR) and Channel Refinement (CR) modules with Non-Local Attention (NLA) for feature enhancement.
• Employed a hierarchical decoder with a Pyramidal Attention Block (PAB) and Convolutional Block Attention Modules (CBAM) for feature fusion and discrimination.

Main Results:

• BiCoMA achieved state-of-the-art performance on five benchmark datasets (Endoscene, ClinicDB, ColonDB, ETIS, Kvasir-SEG).
• The network demonstrated superior generalization capabilities across diverse polyp presentations.
• The proposed architecture maintains practical computational efficiency suitable for real-time clinical applications.

Conclusions:

• The Bilateral Convolutional Multi-Attention Network (BiCoMA) offers a significant advancement in automated colorectal polyp segmentation.
• BiCoMA's hybrid approach effectively addresses challenges related to polyp variability and computational demands.
• This technology holds promise for improving the accuracy and efficiency of colonoscopy-based cancer screening.