Multiscale attention network with structure guidance for colorectal polyp segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate segmentation of colorectal polyps is vital for early diagnosis and treatment planning.
• Existing polyp segmentation methods struggle with variations in polyp size, morphology, and indistinct boundaries.

Purpose of the Study:

• To develop an advanced deep learning framework for precise colorectal polyp segmentation.
• To address the limitations of current methods in handling polyp variability and indistinct structures.

Main Methods:

• Proposed a multiscale attention network with structure guidance (MAN-SG).
• Introduced a structure extraction module (SEM) to capture fine-grained structural information.
• Developed a cross-scale structure guided attention (CSGA) module for effective multiscale feature fusion.
• Utilized Res2Net-50 and PVTv2-B2 as backbone networks.

Main Results:

• MAN-SG demonstrated superior performance on five benchmark polyp segmentation datasets.
• The framework consistently outperformed existing state-of-the-art methods.
• Achieved highly accurate delineation of colorectal polyp structures.

Conclusions:

• The MAN-SG framework offers a highly effective and robust solution for colorectal polyp segmentation.
• The integration of structural guidance through SEM and CSGA modules significantly enhances segmentation accuracy.
• This approach holds promise for improving clinical diagnosis and treatment planning.