Capsule Endoscopy (CE) provides a minimally invasive technique for comprehensive assessment of the gastrointestinal tract. Protruding lesions are frequent findings with heterogeneous prevalence and clinically relevant implications, yet their consistent detection remains challenging due to the extensive duration of CE studies and the influence of reader fatigue and human error. Artificial intelligence has emerged as a promising strategy to strengthen diagnostic accuracy, although robust real-world evidence supporting its clinical value is still limited. This study assessed the diagnostic performance of a deep learning convolutional neural network software for the identification of protruding lesions in complete CE videos under routine clinical conditions, generating clinical evidence that may inform the responsible integration of AI into CE practice and guide future validation efforts.

Our group conducted a prospective, multicenter validation comparing AI-assisted CE interpretation with conventional reading. A total of 423 CE videos, acquired using three different CE systems across ten centers in Portugal, Spain, Brazil, Uruguay, Australia, and the United States, were included. After the initial standard reading, an independent expert performed AI-assisted review using a CNN specifically trained for small bowel protruding lesion detection. Discrepancies between the two readings were evaluated by a third expert from an external center to establish the reference standard. Diagnostic performance was quantified using sensitivity, specificity, and accuracy.

AI-assisted interpretation identified 49 confirmed protruding lesions, compared with 24 detected through conventional reading. Sensitivity was substantially higher with AI (94.2% vs. 46.2%), while specificity was lower (79.5% vs. 93.4%). McNemar’s test demonstrated statistically significant superiority of AI-assisted reading (p = 0.015). Additionally, AI support markedly reduced the mean reading time, with a median examination time of 316 seconds per video.

AI-assisted review of complete CE studies substantially increases the detection of protruding lesions compared with conventional interpretation, accompanied by only a modest reduction in specificity. The inclusion of data from multiple CE systems addresses interoperability limitations, and the participation of centers from six countries across three continents strengthens the generalizability of the findings. Overall, these results support the real-world clinical applicability of AI in CE and provide a rationale for its integration into diagnostic workflows to enhance performance and efficiency