UveAI: Automated fluorescein angiography scoring for retinal inflammation in uveitis

Automated grading of retinal inflammation on fluorescein angiography (FA) in uveitis has historically been limited by the subjectivity, labor intensity, and poor scalability of manual scoring methods. A new study published in Scientific Reports presents UveAI, a modular deep learning framework designed to grade all major retinal inflammatory signs on widefield FA and generate a composite inflammation score aligned with the Angiography Scoring for Uveitis Working Group (ASUWOG) system.¹ The system was developed and validated on the largest clinically annotated FA dataset in uveitis to date.

Background and rationale

Uveitis is the third leading cause of blindness in industrialized countries, with an annual incidence estimated at 17 to 52 cases per 100,000 people.¹ In its most severe forms—intermediate, posterior, and panuveitis—retinal inflammation serves as a central marker of disease activity that informs decisions about initiating and escalating immunomodulatory therapy. Despite FA being the gold standard for assessing retinal inflammation, it is rarely incorporated into clinical trials, which typically rely on visual acuity (VA), macular thickness on optical coherence tomography (OCT), or clinical flare-ups—measures that do not reliably reflect the degree of retinal inflammation visible on FA. The ASUWOG system provides a validated semi-quantitative angiographic scoring framework but remains time-consuming to apply.

Amiot, Pulvirenti, and colleagues, from institutions in Lausanne, Martigny, Grenoble, and Lucerne, extended a previously developed transformer-based system for posterior pole FA grading to enable comprehensive scoring that includes the retinal periphery.¹

Study design and data set

Data were collected retrospectively from patients examined at the Ocular Immune-Infectiology Department at Jules-Gonin Eye Hospital between 2018 and 2021. FA was performed using a novel platform (Spectralis; Heidelberg Engineering), with standard posterior pole imaging and ultra-widefield peripheral quadrant imaging. A total of 644 eyes from 369 patients were included, comprising the training (70%), validation (15%), and test (15%) sets. Images from patients with conditions that could interfere with angiography interpretation—including diabetic retinopathy and vascular occlusions—were excluded.

Expert grading was performed by a senior uveitis specialist using the ASUWOG scoring system across 6 inflammatory signs: macular edema, optic disc hyperfluorescence, vascular leakage in the posterior pole, capillary leakage in the posterior pole, peripheral vascular leakage, and peripheral capillary leakage. An independent test set of 93 eyes was annotated by 3 additional uveitis experts from different institutions to benchmark inter-grader agreement.

UveAI architecture

UveAI integrates 6 individually trained transformer models, each assessing 1 of the 6 inflammatory signs. The system builds on a prior framework developed for posterior pole signs and extends it to peripheral FA images using 2 additional models. For peripheral grading, an automated image selector model—trained on 1,500 manually annotated images—identified peripheral quadrant frames with 98% accuracy. All 6 models used the RETFound Green backbone, a compact retinal foundation model pre-trained using a self-supervised Token Reconstruction strategy. Models were trained using an ordinal cross-entropy loss function, which penalizes larger ordinal classification errors more heavily than smaller ones, reflecting clinical priorities. Grad-CAM saliency maps were used to visualize the regions most influential to model predictions.

Results

The most prevalent inflammatory sign in the cohort was optic disc hyperfluorescence (254 eyes, 39.3%), followed by peripheral vascular leakage (242 eyes, 37.4%). The median ASUWOG score in eyes with active inflammation was 5 (interquartile range, 2 to 11).

Average inter-grader agreement across all 6 signs was 1-OCI = 0.82 (SD = 0.02), accuracy = 0.78 (SD = 0.03) and F1-score = 0.79 (SD = 0.02). Pearson correlation coefficients for the composite ASUWOG score between the 3 independent experts and the senior grader were R = 0.84, 0.86 and 0.82, respectively.

UveAI demonstrated performance on par with or exceeding human inter-grader consistency across all 6 signs. For peripheral vascular leakage, for which no prior automated method existed, the model achieved AUC = 0.95, 1-OCI = 0.90, accuracy = 0.86 and F1-score = 0.85. Peripheral capillary leakage achieved AUC = 0.98, 1-OCI = 0.95, accuracy = 0.91 and F1-score = 0.92. The composite ASUWOG score predicted by UveAI showed a Pearson correlation coefficient of R = 0.96 with the senior expert, with near-unity slope (β = 1.09) and minimal intercept (0.01), comparable to inter-grader agreement metrics. Grad-CAM maps confirmed that models focused on clinically relevant structures in both the posterior pole and periphery. Large errors were mainly observed in eyes with chorioretinal scarring or retinal pigment epithelium (RPE) atrophy.

Limitations

The authors identified several limitations. Certain rare ASUWOG signs—including ischemia and pinpoint leakage—were excluded from model development due to low prevalence in the dataset. The data set, although the largest of its kind in uveitis, remains modest compared with those used for more prevalent conditions. The system was developed and validated on the Heidelberg Spectralis ultra-widefield module; cross-device external validation will be required to confirm robustness across different acquisition platforms. Ground-truth labels were provided by a single senior specialist, which may bias the model toward that grader’s interpretation; the authors note that model agreement with the reference grader was within the range of inter-grader agreement. Prospective clinical validation in larger and more diverse cohorts was identified as a priority for establishing clinical utility.

The authors describe UveAI as the first system to achieve fully automated ASUWOG-aligned grading of both posterior and peripheral inflammatory signs across a large clinical dataset. Future development is planned to incorporate additional imaging modalities, including OCT, and to pursue external validation using data from other institutions.

Reference

1. Amiot V, Pulvirenti R, Jimenez-Del-Toro O, et al. UveAI: clinic-ready scoring of retinal inflammation in uveitis on widefield fluorescein angiography using AI. Sci Rep. doi:10.1038/s41598-026-46069-w