Benchmarking Polyp Segmentation Methods in Narrow-Band Imaging Colonoscopy Images

Area of Science:

• Medical imaging
• Computer vision
• Gastroenterology

Background:

• Deep learning (DL) has advanced polyp segmentation in white-light imaging (WLI) colonoscopy.
• Narrow-band imaging (NBI) enhances polyp visibility but presents segmentation challenges like small sizes, interference, and camouflage.

Purpose of the Study:

• To evaluate the performance of existing DL polyp segmentation methods on NBI colonoscopy images.
• To introduce a new annotated dataset for NBI polyp segmentation.
• To identify areas for improvement in DL models for NBI colonoscopy.

Main Methods:

• A new dataset, PS-NBI2K, with 2,000 annotated NBI colonoscopy images was created.
• 24 recent DL-based polyp segmentation methods were benchmarked on the PS-NBI2K dataset.
• Analyses focused on performance with small polyps, interference, and feature extraction.

Main Results:

• Existing DL methods show limitations in segmenting small polyps and those with significant background interference in NBI images.
• Incorporating both local and global features enhances segmentation performance.
• A trade-off exists between model effectiveness and efficiency, with few methods excelling in both.

Conclusions:

• Current DL polyp segmentation methods require further development for reliable NBI colonoscopy application.
• Future research should focus on robust feature extraction and balancing performance and efficiency.
• The PS-NBI2K dataset release aims to foster advancements in NBI polyp segmentation.