Polyp characterization using deep learning and a publicly accessible polyp video database.

Area of Science:

• Gastroenterology
• Artificial Intelligence
• Medical Imaging

Background:

• Convolutional neural networks (CNNs) for polyp diagnosis often use limited still images.
• Sessile serrated lesions (SSLs) are frequently excluded from training datasets.
• Standard narrow-band imaging (NBI) and NBI-near focus (NBI-NF) are key endoscopic modalities.

Purpose of the Study:

• To develop a CNN using polyp videos for adenoma classification.
• To evaluate CNN performance using standard NBI and NBI-NF.
• To create and share a public polyp video database.

Main Methods:

• Trained a CNN on 16,832 frames from 229 polyp videos, including 56 SSLs.
• Evaluated the CNN on 222 polyp videos across two test sets.
• Benchmarked CNN performance against expert and non-expert endoscopists using a public dataset.

Main Results:

• Achieved high sensitivity (89.7-91.6%) and specificity (88.5-91.9%) for adenoma characterization.
• Demonstrated strong performance for diminutive polyps (sensitivity 87.5-89.9%, specificity 88.2-90.5%).
• CNN performance was comparable to expert endoscopists and superior to non-experts.

Conclusions:

• A single CNN can effectively differentiate adenomas from SSLs and hyperplastic polyps using NBI and NBI-NF.
• The developed CNN meets established performance benchmarks (PIVI-1 and PIVI-2).
• A publicly accessible NBI polyp video database has been established and benchmarked.