AI-Assisted Capsule Endoscopy for Detection of Ulcers and Erosions in Crohn's Disease: A Multicenter Validation Study

Small bowel capsule endoscopy (SBCE) is limited by lengthy, variable interpretation. Artificial intelligence (AI) offers a transformative approach, enabling faster and more accurate lesion detection. This multicenter study aimed to validate an AI model for ulcers and erosions across different SBCE devices.

A multicenter, cross-sectional cohort study was conducted from 2021 to 2024, involving centers in Europe and the United States. Two SBCE devices (PillCamSB3 and Olympus EC-10) were used. The performance of AI-assisted reading, generated by a deep learning model, was compared with standard-of-care (SoC) reading using a reference standard defined by an independent review board. The study utilized 2 SBCE devices (PillCamSB3, Olympus EC-10) and analyzed 259 SBCE exams. The performance of AI-assisted reading generated by the deep learning model was compared with SoC reading against a reference standard defined by an independent review board.

Ulcers and erosions were detected in 93 (35.9%) patients. SoC had 69.6% sensitivity, 99.4% specificity, 98.5% positive predictive value, 85.6% negative predictive value, and 88.8% accuracy. AI-assisted reading detected ulcers and erosions with 90.2% sensitivity, 84.4% specificity, 76.1% positive predictive value, 94.0% negative predictive value, and 86.5% accuracy. The detection yield of AI-assisted reading was superior (P < .001) to conventional SoC reading. The AI-assisted physician SBCE reading identified 568 lesions (94.7%) out of 600 identified by expert board review. The median AI-assisted CE reporting time was 172 seconds per exam.

The AI-assisted SBCE reading achieved superior diagnostic performance compared with SoC, with a substantial decrease in reading time.