Abstract Background Accurate assessment of endoscopic disease activity is essential for the management of Ulcerative colitis (UC). Virtual chromoendoscopy (VCE) enhances mucosal and vascular detail, yet interpretation remains challenging and subject to substantial interobserver variability. Recent AI-models can detect and convert VCE modalities, providing a multimodal and objective evaluation1. However, existing AI systems predominantly analyse isolated frames, lacking spatial context and overlooking patchy inflammation. Frame-based approaches are prone to artefacts and misclassification because they cannot follow mucosal changes across the entire colon. To address these limitations, we developed and externally validated a multiple-instance learning (MIL) model that integrates spatial attention to capture patchy inflammation and mucosal granularity, together with temporal tracking of inflammatory changes across full-length endoscopic videos. Methods Endoscopic videos from high-definition white-light endoscopy(HD-WLE), iScan2, iScan3, and narrow-band imaging(NBI) modalities from UC patients in the LIMUC2, PICaSSO3, and Birmingham (BHM)4 cohorts (1,391, 302, and 54 patients, respectively) were used to develop SpatioMIL. This MIL framework integrates spatial and temporal attention to aggregate frame-level information for robust video-level inflammation detection (Fig 1). Videos were labelled as active inflammation or remission according to Mayo Endoscopic Score. Overall, 784 colonoscopy videos 7,840 frames) from LIMUC dataset were used for model training. External validation was performed on 148 full-length PICaSSO (30,000 frames) and 51 BHM videos (8,000 frames). Model performance in predicting endoscopic remission was evaluated. Results Table 1 details the performance metrics of SpatioMIL across the three cohorts. In the LIMUC cohort, the model achieved a sensitivity, specificity and accuracy of 93.6%, 90.4% and 92.4%, respectively. External validation demonstrated robust generalizability, with excellent performance in predicting endoscopic remission in both PICaSSO and BHM cohorts (sensitivity: 84.0% and 88.0%, specificity: 96.8% and 84.0%, accuracy: 94.6% and 86.9%). Interestingly, SpatioMIL outperformed all state-of-the-art MIL approaches, with 94.6 and 86% accuracy in the PICaSSO and BHM cohort, respectively. Conclusion Our novel SpatioMIL model integrates spatial and temporal attention for UC assessment, enabling accurate disease evaluation by capturing patchy inflammation and mucosal granularity across the entire colon. The model demonstrated excellent performance with strong external validation across diverse VCE modalities, highlighting its potential as a reliable tool for objective and reproducible endoscopic inflammation assessment. References: 1. Iacucci M, Zammarchi I, Santacroce G, et al. A Novel Switching of Artificial Intelligence to Generate Simultaneously Multimodal Images to Assess Inflammation and Predict Outcomes in Ulcerative Colitis-(With Video). Dig Endosc. 2025;37(10):1078-1088. doi:10.1111/den.15067 2. Polat G, Kani HT, Ergenc I, Ozen Alahdab Y, Temizel A, Atug O. Improving the Computer-Aided Estimation of Ulcerative Colitis Severity According to Mayo Endoscopic Score by Using Regression-Based Deep Learning. Inflamm Bowel Dis. 2023;29(9):1431-1439. doi:10.1093/ibd/izac226 3. Iacucci M, Daperno M, Lazarev M, Arsenascu R, Tontini GE, Akinola O, et al. Development and reliability of the new endoscopic virtual chromoendoscopy score: the PICaSSO (Paddington International Virtual ChromoendoScopy ScOre) in ulcerative colitis. Gastrointest Endosc. 2017 Dec;86(6):1118-1127.e5. 4. R. Cannatelli, A. Bazarova, F. Furfaro, et al., Reproducibility of the Electronic Chromoendoscopy PICaSSO Score (Paddington International Virtual ChromoendoScopy ScOre) in Ulcerative Colitis Using Multiple Endoscopic Platforms: A Prospective Multicenter International Study (With Video), Gastrointestinal Endoscopy 96 (2022): 73–83. Conflict of interest: Iacucci, Marietta: Grant: Pentax, Olympus, Eli lilly,Helmsley Personal Fees: Pentax, Pfitzer, Janssen,EliLilly,J & J Pugliano, Cecilia Lina: No conflict of interest Dr. Hughes, Robert: No conflict of interest Lo Bello, Antonio: No conflict of interest Zammarchi, Irene: No conflict of interest Santacroce, Giovanni: No conflict of interest Chaudhari, Ujwala: No conflict of interest Meseguer Esbri, Pablo: No conflict of interest Del Amor, Maria Rocio: No conflict of interest Naranjo, Valery: No conflict of interest Ghosh, Subrata: No conflict of interest Kolawole, Bisi Bode: No conflict of interest Grisan, Enrico: No conflict of interest
Iacucci et al. (Thu,) studied this question.