INTELCAPE: A Deep Learning-Powered System for Automated, High-Accuracy Crohn's Disease Diagnosis via Capsule Endoscopy.

BACKGROUND & AIMS: Capsule endoscopy (CE) is a noninvasive technique for diagnosing Crohn's disease (CD); however, manual interpretation of CE videos is time-consuming and error-prone. We developed an artificial intelligence system, INTELCAPE, to automate CE video analysis for accurate and efficient CD diagnosis. METHODS: This retrospective, multi-center study used data from 2 Chinese hospitals. A multi-task deep learning framework segmented small-intestine regions, detected lesions, and diagnosed CD using CE videos from 757 (Cohort 1) and 115 (Cohort 2) patients. INTELCAPE integrated the ResNet, Transformer, and EfficientNet architectures for hierarchical processing. Performance was benchmarked against clinicians using 3 metrics. This study received Ethics Committee approval (2024ZSLYEC-040). RESULTS: INTELCAPE achieved state-of-the-art performance across all tasks. For small-intestine segmentation, the model showed intersection over union scores of 94.82% (Cohort 1, 95% confidence interval, 93.28%-96.36%) and 96.87% (Cohort 2, 95% confidence interval, 94.63%-99.12%). For lesion detection, it achieved area under the curve values of 0.993 (Cohort 1) and 0.980 (Cohort 2), with 99.33% classification accuracy, which was comparable to that of specialists (97.83%) but superior to that of residents (91.05%; P < .001). For CD diagnosis, INTELCAPE demonstrated robust generalizability, achieving area under the curve values of 0.982 (Cohort 1) and 0.984 (Cohort 2) with 90% diagnostic accuracy, comparable to that of specialists (93.33%) but 10-fold faster (P < .001). INTELCAPE improved doctors' diagnostic accuracy (76.7%-94.8%; P < .001), while reducing their interpretation time (67.9-22.5 minutes; P < .001). CONCLUSIONS: INTELCAPE improved CD diagnosis by automating CE video analysis, thereby enhancing accuracy and efficiency, particularly for less-experienced clinicians.