%Aigaion2 BibTeX export from Idiap Publications
%Friday 29 May 2026 03:39:17 PM

@ARTICLE{Amiot_SCIREP_2026,
                      author = {Amiot, Victor and Pulvirenti, Roberto and Jimenez-del-Toro, Oscar and Ott, Muriel and Bogaciu, Teodora-Elena and Banerjee, Shalini and Amstutz, Christoph and Odobez, Jean-Marc and Chiquet, Christophe and Guex-Crosier, Yan and Bergin, Ciara and Meloni, Ilenia and Anjos, Andr{\'{e}} and Hoogewoud, Florence and Tomasoni, Mattia},
                    keywords = {Clinical translation, deep learning, disease grading, fluorescein angiography, inter-grader agreement, Uveitis, vasculitis},
         mainresearchprogram = {AI for Life},
                       month = may,
                       title = {UveAI: clinic-ready scoring of retinal inflammation in uveitis on widefield fluorescein angiography using AI},
                     journal = {Scientific reports},
                        year = {2026},
                         url = {https://www.nature.com/articles/s41598-026-46069-w},
                         doi = {10.1038/s41598-026-46069-w},
                    abstract = {Retinal inflammation is a key determinant of visual prognosis in uveitis, yet its assessment on fluorescein angiography remains subjective, labor-intensive, and insufficiently scalable for clinical trials or large cohort studies. Fluorescein angiography is the gold standard for assessing retinal inflammation. However, its scoring remains challenging, as the process is complex and time-consuming, limiting routine use in clinical trials and patient care. We present UveAI, a modular deep learning framework that grades all major retinal inflammatory signs in fluorescein angiography across posterior pole and periphery to generate an ASUWOG-aligned inflammation score. Trained on 3,220 FA images from 644 eyes (369 patients), UveAI integrates six transformer models detecting macular edema, optic disc hyperfluorescence, and vascular and capillary leakage in the posterior pole and periphery. On an independent test set, UveAI showed high concordance with an expert grader for total score (R = 0.96) and strong performance for individual signs (mean AUC = 0.952). Grad-CAM maps confirmed clinically relevant focus, supporting automated, standardised FA scoring in uveitis.}
}