GAR-Net: Guided Attention Residual Network for Polyp Segmentation from Colonoscopy Video Frames

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Colorectal cancer is a leading cause of cancer deaths, with polyps being its precursor.
• Accurate polyp detection during colonoscopy is crucial for early diagnosis and cancer prevention.
• Existing segmentation methods struggle with irregular or small polyps, leading to inaccurate delineation.

Purpose of the Study:

• To develop an end-to-end pixel-wise polyp segmentation model for improved accuracy in colonoscopy.
• To introduce a novel Guided Attention Residual Network (GAR-Net) that refines segmentation maps.
• To address limitations in current models regarding irregular and small polyp segmentation.

Main Methods:

• Proposed an end-to-end pixel-wise polyp segmentation model, Guided Attention Residual Network (GAR-Net).
• Integrated enhanced residual blocks to suppress noise and capture low-level features.
• Introduced Guided Attention Learning, a novel attention mechanism for refined attention maps across layers.

Main Results:

• GAR-Net demonstrated superior performance compared to FCN8, SegNet, U-Net, ResUNet, and DeepLabv3.
• Achieved a 91% Dice coefficient and 83.12% mIoU on the CVC-ClinicDB dataset.
• Attained an 89.15% Dice coefficient and 81.58% mIoU on the Kvasir-SEG dataset.

Conclusions:

• GAR-Net provides a robust solution for accurate polyp segmentation from colonoscopy video frames.
• The model's ability to capture refined attention maps enhances segmentation accuracy for polyps of varying shapes and sizes.
• This advanced segmentation technique can significantly aid endoscopists in detecting polyps and preventing colorectal cancer progression.