S3085 Multicenter Real-World Validation of an AI-Based System for Detecting Protruding Lesions in Capsule Endoscopy

Introduction: Capsule endoscopy (CE) is a non-invasive imaging method widely used for small bowel evaluation. Among its findings, protruding lesions are relatively common but challenging to detect due to the time-consuming nature of CE review and susceptibility to human oversight. Artificial intelligence (AI) tools, particularly convolutional neural networks (CNNs), may improve diagnostic performance. However, robust real-world data validating their effectiveness remain limited. This study aimed to evaluate the diagnostic accuracy of an AI system for detecting small bowel protruding lesions in full-length CE videos across diverse clinical environments. Methods: In a prospective, multicenter study, 398 CE videos from 9 centers across Portugal, Spain, Brazil, Australia, and the United States were analyzed. The dataset included videos from 3 different CE systems. All exams underwent conventional reading followed by an AI-assisted interpretation using a CNN trained to detect protruding lesions. An independent expert adjudicated discrepancies between the 2 readings, establishing the reference standard. Performance was assessed using sensitivity, specificity, and overall accuracy. Mean reading time per video was also recorded. Results: The AI system identified 107 confirmed protruding lesions, compared to 47 detected via standard reading. Sensitivity with AI was significantly higher (92.8% vs 46.3%), while specificity was moderately lower (79.8% vs 93.5%). The AI model’s superiority over conventional interpretation was statistically significant (P < 0.001). Additionally, AI-assisted reading reduced the average interpretation time to 316 seconds per video. Conclusion: This multicenter study demonstrates that AI-assisted capsule endoscopy substantially enhances detection of protruding lesions, nearly doubling sensitivity compared to conventional reading, with an acceptable trade-off in specificity. The inclusion of data from multiple CE platforms and international centers supports the generalizability and interoperability of the model. These findings highlight the potential of AI to streamline CE workflows and improve diagnostic accuracy in routine clinical practice.