Introduction
Small bowel capsule endoscopy (SBCE) is a key diagnostic tool for evaluating obscure gastrointestinal bleeding and suspected inflammatory bowel disease (IBD). Artificial intelligence (AI) has been proposed to improve efficiency, standardization, and diagnostic accuracy. However, comparative evidence with experienced human readers in real-world settings remains limited.
Aims & Methods
This study aimed to compare the diagnostic performance of a validated AI system with that of experienced gastroenterologists in detecting clinically significant lesions on SBCE.
A retrospective single-center study was conducted, including patients who underwent SBCE using OMOM® HD capsules between January and September 2025. Readings were performed independently by experienced gastroenterologists and a validated AI system. Only clinically significant small bowel lesions, defined independently by expert reviewers, were considered. Sensitivity for lesion detection was calculated for each method, using the combined total of all identified lesions as the reference standard. Statistical comparisons were performed using McNemar’s test.
Results
A total of 62 patients were included (mean age 52.5±18.5 years; 54.8% female). The main indication for SBCE was suspected small bowel bleeding (48.4%). Overall diagnostic agreement between human readers and AI was 100%. In total, 247 clinically relevant lesions were identified. Overall sensitivity was identical for gastroenterologists and the AI system (97.2% [95% CI 93.9–98.8] for both methods). For ulcer/erosion detection (n=199), sensitivity was 96.5% for human readers and 97.5% for AI (p=0.564). For angiectasias (n=32), sensitivity was 100% versus 93.8% (p=0.5), with no statistically significant differences.
Conclusion
In real-world clinical practice, AI demonstrated diagnostic performance equivalent to that of experienced gastroenterologists, with comparable sensitivity for detecting clinically significant lesions on SBCE. These results demonstrate that AI can reliably support clinical decision-making and safely integrate into SBCE workflows.