Objectives
Small bowel capsule endoscopy (CE) produces lengthy videos that are time-consuming to review and susceptible to missed lesions. We evaluated whether an open-source, pretrained artificial intelligence (AI) model (SEE-AI) could improve diagnostic performance and interpretation efficiency compared with conventional reading.
Methods
We retrospectively analyzed 249 PillCam SB3 examinations performed between 2007 and 2022 at six hospitals, using a two-reader crossover design. SEE-AI (confidence threshold 0.1) generated annotated videos with bounding boxes for eight lesion categories. The primary endpoints were sensitivity for lesion detection on a per-lesion and per-patient basis. Secondary endpoints included specificity, predictive values, overall accuracy, and reading time. A prespecified subgroup analysis evaluated cases of suspected small-bowel bleeding (SSBB), focusing on Saurin P1+P2 hemorrhagic lesions.
Results
Across 1550 adjudicated lesions, AI-assisted reading demonstrated higher sensitivity than conventional reading (per-lesion: 98.8% [1532/1550] vs. 86.4% [1339/1550]; per-patient: 99.1% [464/468] vs. 80.3% [376/468]; both p < 0.0001). The mean reading time decreased from 17.9 to 13.7 min (p < 0.0001). In SSBB cases (n = 131), sensitivity for P1+P2 lesions improved on both a per-lesion basis (98.2% [439/447] vs. 82.8% [370/447]) and per-patient basis (98.6% [145/147] vs. 73.5% [108/147]), with a shorter reading time (14.1 vs. 18.0 min; all p < 0.0001).
Conclusions
In this multicenter evaluation, SEE-AI significantly improved lesion detection and reduced reading time for CE interpretation, including SSBB cases, while maintaining openness and reproducibility. AI-assisted reading may reduce clinicians‘ workload and support the adoption of SEE-AI as a practical tool – and a potential future standard of care – for small bowel CE.
Trial registration
N/A.