Gastroenterologist-level Identification of Small Bowel Diseases and Normal Variants by Capsule Endoscopy Using a Deep-learning Model

Abstract Background & Aims Capsule endoscopy has revolutionized investigation of the small bowel. However, this technique produces a video that is 8–10 hours long, so analysis is time consuming for gastroenterologists. Deep convolutional neural networks (CNNs) can recognize specific images among a large variety. We aimed to develop a CNN-based algorithm to assist in evaluation of small bowel capsule endoscopy (SB-CE) images. Methods We collected 113,426,569 images from 6970 patients who underwent SB-CE at 77 medical centers from July 2016 through July 2018. A CNN-based auxiliary reading model was trained to differentiate abnormal from normal images using 158,235 SB-CE images from 1970 patients. Images were categorized as normal, inflammation, ulcer, polyps, lymphangiectasia, bleeding, vascular disease, protruding lesion, lymphatic follicular hyperplasia, diverticulum, parasite, and other. The model was further validated in 5000 patients (no patient was overlap with 1970 patients in training set); the same patients were evaluated by conventional analysis and CNN-based auxiliary analysis by 20 gastroenterologists. If there was agreement in image categorization between the conventional analysis and CNN-model, no further evaluation was performed. If there was disagreement between the conventional analysis and CNN-model, the gastroenterologists re-evaluated the image to confirm or reject the CNN categorization. Results In the SB-CE images from the validation set, 4206 abnormalities were identified after final consensus evaluation, in 3280 patients. The CNN-based auxiliary model identified abnormalities with 99.88% sensitivity in the per-patient analysis (95% CI, 99.67–99.96) and 99.90% sensitivity in the per-lesion analysis (95% CI, 99.74–99.97). Conventional reading by the gastroenterologists identified abnormalities with 74.57% sensitivity (95% CI, 73.05–76.03) in the per-patient analysis and 76.89% in the per-lesion analysis (95% CI, 75.59–78.15). The mean reading time per patient was 96.6±22.53 min by conventional reading and 5.9±2.23 min by CNN-based auxiliary reading ( P Conclusions We validated the ability of a CNN-based algorithm to identify abnormalities in SB-CE images. The CNN-based auxiliary model identified abnormalities with higher levels of sensitivity and significantly shorter reading times than conventional analysis by gastroenterologists. This algorithm provides an important tool to help gastroenterologists analyze SB-CE images more efficiently and more accurately.