Document resource
Introduction Artificial intelligence (AI) is increasingly integrated into capsule endoscopy (CE), with growing evidence supporting its role in enhancing pathology recognition and reducing physician reading time. 1 There is paucity of studies assessing its role in training. This study evaluates the ability of TOP100, an integrated software that selects 100 images most likely to contain abnormal pathology, in supporting CE readers with varying levels of training, by comparing AI assisted trainees interpretation with expert assessment.Methods We conducted a single centre retrospective cohort study of patients who underwent small bowel CE using the PillCam™ SB3 system in 2025. Two expert readers, who had previously read the full CE videos, performed a blinded review of the same CE previously read using TOP100 only. The same cases were independently reviewed by a senior trainee (ST) with capsule and endoscopy experience, 2 junior trainees (JT) independent in endoscopy and minimal CE experience (<10), and a medical student who received one hour of basic pathology recognition training. Predefined abnormalities being noted include erosions, ulcers, active bleeding, tumours, bulges, atrophy, and angioectasia. The agreement between trainee and expert for each finding category was analysed.Results A total of 50 CE videos were reviewed by each reader. Indications were overt bleeding (12.2%), iron deficiency anaemia (40.8%) and inflammatory bowel disease (40.8%). All videos were complete, 3 (6.1%) had inadequate bowel preparation. Ulcers and angioectasia demonstrated the highest concordance, with the greatest agreement observed for the ST (κ 0.912 and 0.703, respectively; p<0.001), and moderate to substantial agreement for the JTs and medical student (κ 0.457-0.703; p<0.001). Active bleeding was detected reliably across all readers, with the highest agreement seen among JT2 followed by the medical student then ST and JT1 (κ 0.811, 0.669 and 0.558, respectively; p<0.001). Agreement for erosions was more variable with the ST having the highest agreement with experts. In contrast, agreement for atrophy and bulges was poorer across all trainees. Mean reading time using TOP100 varied by experience: experts took 1m48s per CE, ST 1m15s, JT1 2m37s, JT2 6m36s and medical student 4m11s.Conclusion The TOP100 demonstrated good concordance with expert assessment for key small bowel pathologies, particularly ulcers, angioectasia and active bleeding. Agreement was highest among the senior trainee but remained meaningful for junior trainees and medical student, suggesting a role in supporting pathology recognition and improving efficiency across training levels. Further prospective studies with larger and more diverse trainee cohorts are required to confirm these findings.Reference Ho JCL, Qian Z, Lau LHS, et al. Artificial intelligence in digestive endoscopy training-the past, present, and future. Dig Endosc. 2026;38:e70047.