Background Further development of deep learning-based artificial intelligence (AI) technology to automatically diagnose multiple abnormalities in small-bowel capsule endoscopy (SBCE) videos is necessary. We aimed to develop an AI model, to compare its diagnostic performance with doctors of different experience levels, and to further evaluate its auxiliary role for doctors in diagnosing multiple abnormalities in SBCE videos.
Methods The AI model was trained using 280 426 images from 2565 patients, and the diagnostic performance was validated in 240 videos.
Results The sensitivity of the AI model for red spots, inflammation, blood content, vascular lesions, protruding lesions, parasites, diverticulum, and normal variants was 97.8 %, 96.1 %, 96.1 %, 94.7 %, 95.6 %, 100 %, 100 %, and 96.4 %, respectively. The specificity was 86.0 %, 75.3 %, 87.3 %, 77.8 %, 67.7 %, 97.5 %, 91.2 %, and 81.3 %, respectively. The accuracy was 95.0 %, 88.8 %, 89.2 %, 79.2 %, 80.8 %, 97.5 %, 91.3 %, and 93.3 %, respectively. For junior doctors, the assistance of the AI model increased the overall accuracy from 85.5 % to 97.9 % (P < 0.001, Bonferroni corrected), comparable to that of experts (96.6 %, P > 0.0125, Bonferroni corrected).
Conclusions This well-trained AI diagnostic model automatically diagnosed multiple small-bowel abnormalities simultaneously based on video-level recognition, with potential as an excellent auxiliary system for less-experienced endoscopists.