Training a computer-aided polyp detection system to detect sessile serrated adenomas using public domain colonoscopy videos

• Main Authors: Taibo Li, Jeremy R. Glissen Brown, Kelovoulos Tsourides, Nadim Mahmud, Jonah M. Cohen, Tyler M. Berzin
• Format: Article
• Language: English
• Published: Georg Thieme Verlag KG 2020-10-01
• Series: Endoscopy International Open
• Online Access: http://www.thieme-connect.de/DOI/DOI?10.1055/a-1229-3927
• ISSN: 2364-3722; 2196-9736

Background Colorectal cancer (CRC) is a major public health burden worldwide, and colonoscopy is the most commonly used CRC screening tool. Still, there is variability in adenoma detection rate (ADR) among endoscopists. Recent studies have reported improved ADR using deep learning models trained on videos curated largely from private in-house datasets. Few have focused on the detection of sessile serrated adenomas (SSAs), which are the most challenging target clinically. Methods We identified 23 colonoscopy videos available in the public domain and for which pathology data were provided, totaling 390 minutes of footage. Expert endoscopists annotated segments of video with adenomatous polyps, from which we captured 509 polyp-positive and 6,875 polyp-free frames. Via data augmentation, we generated 15,270 adenomatous polyp-positive images, of which 2,310 were SSAs, and 20,625 polyp-negative images. We used the CNN AlexNet and fine-tuned its parameters using 90 % of the images, before testing its performance on the remaining 10 % of images unseen by the model. Results We trained the model on 32,305 images and tested performance on 3,590 images with the same proportion of SSA, non-SSA polyp-positive, and polyp-negative images. The overall accuracy of the model was 0.86, with a sensitivity of 0.73 and a specificity of 0.96. Positive predictive value was 0.93 and negative predictive value was 0.96. The area under the curve was 0.94. SSAs were detected in 93 % of SSA-positive images. Conclusions Using a relatively small set of publicly-available colonoscopy data, we obtained sizable training and validation sets of endoscopic images using data augmentation, and achieved an excellent performance in adenomatous polyp detection.