Introduction
Barrier healing is emerging as a key therapeutic target in Inflammatory bowel disease (IBD), and is increasingly associated with improved long-term outcomes than endoscopic and histological remission alone. However, objective and reproducible assessment of barrier integrity remains challenging. Probe-based confocal laser endomicroscopy (pCLE) enables real-time structural and functional assessment of epithelial and vascular barrier features, including villi and crypt architecture, goblet cells, fluorescein leakage and vessel architecture. However, its interpretation remains complex and highly operator-dependent, limiting reproducibility and widespread application. 1. We aimed to develop and validate an artificial intelligence (AI)-driven pCLE system for automated standardised assessment of epithelial and vascular barrier impairment (BI) in IBD.
Aims & Methods
A total of 297 high-quality pCLE videos from IBD patients undergoing colonoscopy in two tertiary centers (University College Cork, Ireland and University of Calgary, Canada) were considered. Experienced endoscopists assessed the presence of overall barrier and feature-specific impairment according to a previously developed pCLE-scoring system 2, providing a reference standard for AI. 196 videos from the Calgary cohort were used to train and validate the framework (using cross-validation) and 101 videos from the Cork cohort were considered for external validation.
An AI algorithm was developed to predict pCLE-based overall- and feature-level BI using a weakly supervised learning paradigm that exclusively employs video-level scoring. First, frame-level representations were extracted with a DINOv2 foundation model 3. Later, a multiple instance learning (MIL) model aggregates frame-features into a global representation to produce the predicted score of the video4. The model’s diagnostic performance to predict overall, epithelial, vascular and feature-level impairment was evaluated.
Results
Overall, 297 videos from 141 IBD patients were considered. BI was detected in 187/297 (63.0%) videos, including 68/101 (67.3%) in Cork and 119/196 (60.7%) in the Calgary cohort.
Table 1 details the diagnostic performance of the model. The AI algorithm achieved remarkable performance in both the development and validation cohorts, with 71.4% (95% 63.3-79.5) and 89.7% (95% CI 82.3-97.1) sensitivity, 71.4% (95% CI 61.3-81.5) and 39.4% (95% CI 22.7-56.1) specificity, 71.4% (95% CI 65.1-77.8) and 73.3% (95% CI 64.6-81.9) accuracy, and 0.77 and 0.72 AUC for detecting overall BI at the video level. Performance remained consistent for most epithelial and vascular components, although slightly lower for villi/crypts architecture and goblet cells, likely reflecting the more challenging assessment of these features with pCLE.
| Overall BI | Epithelial BI | Vascular BI | Villi/ crypts architecture | Goblet cells | Vessel architecture | Fluorescein leakage |
| Development cohort (N=196) |
AUC % (95% CI) | 0.77 | 0.74 | 0.77 | 0.80 | 0.78 | 0.78 | 0.77 |
F1 score % (95% CI) | 75.2 (68.8-81.6) | 74.2 (67.5-80.9) | 75.3 (68.7-81.9) | 68.6 (59.0-78.0) | 76.5 (70.0-83.0) | 63.5 (54.1-72.7) | 76.5 (70.2-82.8) |
Accuracy % (95% CI) | 71.4 (65.1-77.8) | 70.9 (64.4-77.3) | 70.9 (64.4-77.3) | 78.1 (72.3-83.9) | 74.0 (67.9-80.1) | 68.9 (62.4-75.4) | 72.5 (66.2-78.6) |
| Validation cohort (N=101) |
AUC % (95% CI) | 0.72 | 0.71 | 0.75 | 0.58 | 0.73 | 0.79 | 0.74 |
F1 score % (95% CI) | 81.9 (72.4-89.6) | 69.6 (60.0-79.2) | 79.2 (70.9-87.3) | 47.9 (35.2-60.6) | 53.1 (42.1-64.1) | 75.9 (66.2-85.6) | 77.7 (69.8-85.6) |
Accuracy % (95% CI) | 73.3 (64.6-81.9) | 59.4 (49.8-68.9) | 70.3 (61.4-79.2) | 50.5 (40.8-60.2) | 47.5 (42.1-64.1) | 72.3 (63.6-81.0) | 69.3 (60.4-78.2) |
Table. Diagnostic performance of the AI algorithm to predict barrier impairment on pCLE videos.
Conclusion
Our AI-driven pCLE enables automated, standardised in vivo assessment of epithelial and vascular barrier features, offering an objective, reproducible tool to monitor barrier healing as a novel therapeutic endpoint in IBD.
References
1. Iacucci M, Jeffery L, Acharjee A, et al. Computer-Aided Imaging Analysis of Probe-Based Confocal Laser Endomicroscopy With Molecular Labeling and Gene Expression Identifies Markers of Response to Biological Therapy in IBD Patients: The Endo-Omics Study. Inflamm Bowel Dis. 2023;29(9):1409-1420. doi:10.1093/ibd/izac233
2. Iacucci M, Majumder S, Zammarchi I, et al. Automated real-time imaging of intestinal barrier integrity and molecular profiling for early outcome prediction in inflammatory bowel disease - Endo-Histo-Barrier-Omics study. J Crohns Colitis.
3. Oquab M, Darcet T, Moutakanni T, et al. DINOv2: Learning Robust Visual Features without Supervision. Published online February 2, 2024.
4. Carretero I, Meseguer P, et al. Time-aware semantic prototypes for weakly-supervised endomicroscopy video classification. Published online April, 2026.