Frequency-based boundary-guided attention network for domain generalizable polyp segmentation from colonoscopy...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate polyp segmentation in colonoscopy is vital for colorectal cancer prevention and diagnosis.
• Challenges include poor boundary details due to reflections, complex shapes, and low contrast, hindering performance in unobserved clinical settings.

Purpose of the Study:

• To develop a domain-generalizable Frequency-based Boundary-guided Attention Network (FBGANet) for robust polyp segmentation.
• To enhance polyp detection accuracy across various clinical domains, including unseen ones.

Main Methods:

• Proposed FBGANet utilizes a DCT-based component decomposition module (DCT CDM) to remove noise from out-domain data.
• Incorporated a Boundary-guided Attention Block (BGA Block) to preserve high-frequency boundary details for precise segmentation.

Main Results:

• FBGANet demonstrated superior domain generalization compared to state-of-the-art methods.
• Achieved higher Dice Score Coefficient (DSC) and mIoU on both in-domain and out-domain datasets.
• Maintained a reasonable inference speed of 0.035 s/image.

Conclusions:

• FBGANet enables precise polyp detection and segmentation in colonoscopy images, even in challenging, unobserved clinical settings.
• The model's domain generalizability improves diagnostic confidence and has the potential to advance robotic-assisted healthcare and patient outcomes.