LightCF-Net: A Lightweight Long-Range Context Fusion Network for Real-Time Polyp Segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Colorectal cancer diagnosis relies on accurate polyp segmentation from colonoscopy videos.
• Existing methods, especially Transformer-based ones, face challenges with real-time performance due to high computational load.
• There is a need for efficient and accurate automatic polyp segmentation techniques for clinical applications.

Purpose of the Study:

• To propose a novel lightweight network, LightCF-Net, for real-time automatic polyp segmentation.
• To improve the distinction between polyps and background noise for enhanced segmentation accuracy.
• To maintain real-time performance while modeling long-range spatial dependencies.

Main Methods:

• Developed LightCF-Net, a lightweight long-range context fusion network.
• Introduced a Fusion Attention Encoder (FAEncoder) integrating Large Kernel Attention (LKA) and channel attention.
• Incorporated a Visual Attention Mamba (VAM) module in skip connections and a Pyramid Split Attention (PSA) module in the bottleneck layer.

Main Results:

• LightCF-Net demonstrated superior segmentation accuracy compared to state-of-the-art lightweight methods.
• The proposed network achieved higher accuracy with reduced processing time.
• Consistent performance was observed across four benchmark datasets: Kvasir-SEG, CVC-ClinicDB, BKAI-IGH, and ETIS.

Conclusions:

• LightCF-Net offers an effective solution for real-time automatic polyp segmentation in colonoscopy.
• The network's architecture successfully balances accuracy and computational efficiency.
• The proposed method shows significant potential for computer-assisted diagnosis of colorectal cancer.