HCA-Net: Hierarchical Contextual Attention Network for Lightweight and Accurate Polyp Segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Early detection of colorectal polyps is vital for cancer prevention.
• Automatic segmentation of polyps in colonoscopy images is challenging due to low contrast, blurred boundaries, and scale variations.
• Existing segmentation networks often suffer from feature redundancy and semantic inconsistency, limiting accuracy and efficiency.

Purpose of the Study:

• To develop a lightweight and efficient deep learning framework for accurate automatic segmentation of colorectal polyps.
• To address the limitations of existing methods in handling low-contrast and boundary-ambiguous colonoscopy images.
• To improve the feasibility of real-time clinical application for polyp segmentation.

Main Methods:

• Proposed Hierarchical Contextual Attention Network (HCA-Net) with Redundancy-Suppressed Dual-Path Downsampling (RS-DPD) and Boundary-Aware Semantic Alignment Upsampling (BA-SAU) modules.
• Developed a Hierarchical Contextual Attention (HCA) mechanism for efficient global modeling and local boundary restoration.
• Introduced a composite boundary-aware loss function to enhance pixel-level accuracy and structural consistency.

Main Results:

• HCA-Net achieved state-of-the-art (SOTA) segmentation accuracy on public colorectal polyp datasets.
• Demonstrated significantly improved efficiency compared to existing complex models.
• Maintained robustness in segmenting polyps under low-contrast and blurred-boundary conditions.

Conclusions:

• HCA-Net offers a promising solution for accurate and efficient polyp segmentation in clinical screening.
• The proposed framework effectively handles challenging image conditions, enhancing early cancer detection capabilities.
• The lightweight design makes HCA-Net suitable for real-time applications in colonoscopy.