Weakly Supervised Polyp Segmentation from an Attention Receptive Field Mechanism.

Area of Science:

• Medical Imaging
• Computational Pathology
• Gastroenterology

Background:

• Colorectal cancer is a leading global cancer, with polyps as key biomarkers.
• Colonoscopies are crucial for polyp detection, but have a miss rate up to 26% and are subject to expert bias.
• Existing computational methods for polyp segmentation require expert-labeled data and controlled scenarios.

Purpose of the Study:

• To develop a robust polyp segmentation method for colonoscopy images.
• To overcome limitations of supervised learning by incorporating weakly supervised learning.
• To reduce diagnostic errors caused by expert bias and polyp miss rates.

Main Methods:

• An attention receptive field mechanism was employed to learn non-local pixel relationships for polyp shape recovery.
• A weakly supervised learning scheme was utilized, including unlabeled background frames to differentiate polyps from intestinal folds.
• The model was trained and evaluated on public datasets, including CVC-Colon DB.

Main Results:

• The proposed method achieved 95.1% precision on the CVC-Colon DB dataset.
• The approach demonstrated competitive performance on other datasets.
• The attention mechanism robustly recovered polyp shapes, outperforming state-of-the-art methods.

Conclusions:

• The developed strategy offers a novel approach to support polyp segmentation tools in clinical colonoscopy routines.
• This method effectively handles redundant background information in colonoscopy sequences.
• The findings suggest improved accuracy and reliability in automated polyp detection during colonoscopies.