Intestinal barrier healing is an emerging therapeutic target in Inflammatory Bowel Disease (IBD), associated with deep mucosal healing and improved long-term outcomes (1). Ultra-magnification Endocytoscopy (ECS) enables real-time, in vivo cellular-level assessment of the intestinal barrier, with strong histological correlation. However, its clinical application remains limited to expert centers due to the complexity of image interpretation (2). We aimed to develop an artificial intelligence (AI)- driven system to standardise ECS-based evaluation of barrier integrity and predict adverse outcomes.

Thirty-five high-quality ECS videos (2636 frames for crypt architecture, 2596 for goblet cells, 1009 for vascular pattern, and 702 for villi architecture) from Ulcerative colitis (UC) and Crohn’s disease (CD) patients were considered. Using five-fold cross validation, the frames from 28 videos were used for training the model, while the remaining were used for testing. Expert endoscopists scored barrier impairment according to a previously validated ECS (1), providing a reference standard for AI training. Separate convolutional neural networks (ResNet50, pretrained GastroNet data) were trained to identify abnormalities in villi/crypt architecture, vascular structure, and goblet cells density. The scores from the model training at frame level were aggregated to obtain a score at video level for each feature (3,4). A logistic regression model predicted overall barrier impairment, as well as major clinical outcomes (Fig 1). Diagnostic performance metrics were expressed as sensitivity, specificity, accuracy, positive and negative predictive values, and area under the curve (AUC).  

Thirty-five patients (19 UC and 16 CD) were included. Our novel AI algorithm achieved high diagnostic accuracy in detecting overall and features-specific barrier impairment (Tab 1). For overall barrier impairment, the model yielded a sensitivity of 83%, specificity of 100%, accuracy of 97%, and an AUC of 98%. Feature-specific assessment at the video level also demonstrated strong metrics performance across epithelial and vascular features. The model accurately identified abnormalities in crypt architecture (91% AUC), goblet cells (90% AUC), villi architecture (94% AUC), and vascular pattern (96% AUC). Notably, clinical outcome prediction at 6 months achieved an accuracy of 89% for UC, whereas it was only 44% for CD. 

Our AI-driven ECS model provided the first automated and standardised assessment of intestinal barrier integrity in IBD. By overcoming the subjectivity and complexity of current evaluation methods, it enables barrier healing to emerge as a measurable and therapeutic endpoint in clinical practice and clinical trials.