Development and validation of a deep-learning algorithm for the detection of polyps during colonoscopy

Area of Science:

• Gastroenterology
• Artificial Intelligence
• Medical Imaging

Background:

• Colonoscopy is crucial for colon cancer prevention through polyp detection and removal.
• Adenomatous polyp detection rates vary significantly among endoscopists.
• Improving real-time polyp detection during colonoscopy is a key clinical challenge.

Purpose of the Study:

• To develop and validate a deep-learning algorithm for real-time polyp detection during colonoscopy.
• To assess the algorithm's sensitivity and specificity in identifying polyps.
• To evaluate the algorithm's potential to assist endoscopists and assess performance.

Main Methods:

• A deep-learning algorithm was developed using data from 1,290 patients.
• The algorithm was validated on 27,113 colonoscopy images from 1,138 patients.
• Performance was further tested on public datasets and colonoscopy videos, assessing per-image and per-polyp sensitivity and specificity.

Main Results:

• The algorithm achieved high per-image sensitivity (94.38%) and specificity (95.92%) on clinical data.
• Validation on a public database showed 88.24% per-image sensitivity.
• Real-time analysis processed at least 25 frames per second with low latency (76.80 ± 5.60 ms).

Conclusions:

• A machine-learning algorithm demonstrates high accuracy for real-time polyp detection in colonoscopies.
• The AI tool has the potential to aid endoscopists, improving polyp detection and adenoma detection rates.
• This technology could help standardize and assess endoscopist performance in colon polyp detection.