BAASNet: boundary-aware deep learning for accurate polyp segmentation in colonoscopy

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Gastroenterology

Background:

• Colorectal polyps are crucial for early cancer detection via colonoscopy.
• Accurate polyp segmentation aids treatment planning and diagnosis.
• Existing deep learning models struggle with image noise, texture, and boundary variations.

Purpose of the Study:

• Introduce BAASNet, a Boundary-Aware Attention-Based Segmentation framework for polyp segmentation.
• Address limitations of current models, including manual annotation burdens.
• Enhance automated polyp detection and segmentation in colonoscopy.

Main Methods:

• Developed BAASNet, incorporating a boundary-aware loss function for improved edge delineation.
• Evaluated the model on nine diverse public datasets across five imaging modalities.
• Tested on two center-wise polyp detection benchmarks to assess generalization.

Main Results:

• BAASNet demonstrated strong generalization capabilities across multiple datasets and modalities.
• Achieved a mean Dice similarity coefficient (mDSC) of at least [FORMULA: SEE TEXT] on PolypDB across all modalities.
• Showcased an average absolute improvement of approximately [FORMULA: SEE TEXT] in Dice across benchmarks, with relative gains from [FORMULA: SEE TEXT] to [FORMULA: SEE TEXT].

Conclusions:

• BAASNet significantly improves polyp segmentation accuracy and robustness.
• The model shows potential for real-time clinical deployment in automated colonoscopy.
• Highlights the effectiveness of boundary-aware mechanisms in deep learning for medical image segmentation.