Background/Objectives: Capsule endoscopy (CE) is a non-invasive tool for evaluating small bowel pathology, but prolonged reading times can lead to reader fatigue. Artificial intelligence-integrated reading software has shown promise to overcome this. This study assesses the diagnostic performance of TOP100 against standard human reading (SR) across different clinical indications. Methods: A retrospective single-centre cohort study was conducted at a tertiary referral centre, including patients who underwent PillCam™SB3 CE between January 2023 to August 2025. Initial reading was performed by expert CE readers (>1000 CE experience) considered as SR. Second blinded reading was performed with TOP100. CE findings from SR were compared with those from TOP100. Results: A total of 1382 CE were identified, of which 1374 had complete examinations and TOP100 data for analysis. The most common indications were inflammatory bowel disease (IBD) (51.5%) and iron-deficiency anaemia (IDA) (25.0%). Diagnostic yield was significantly higher with SR than TOP100 (36.0% vs. 27.5%, p < 0.001). TOP100 demonstrated moderate sensitivity but high specificity; 64.0% and 99.0% for active bleeding, 58.2% and 97.5% for angiodysplasia. In overt bleeding and IDA, sensitivity, and specificity for P2 lesions were 61.0% and 94.3%, respectively. Overall sensitivity and specificity were 55.3% and 95.9% for ulcers, 43.1% and 89.0% for erosions. In IBD, ulcer detection was similar, while TOP100 sensitivity for erosions improved by 9.4%. Conclusions: SR remains the reference standard for reading and reporting capsule endoscopies. However, TOP100 may be a useful adjunct to support interpretation and case prioritisation, especially in high-volume centres.
Devanand et al. (Sat,) studied this question.