Fully Automatic Grading of Retinal Vasculitis on Fluorescein Angiography Time-lapse from Real-world Data in Clinical Settings

The objective of this study is to showcase a pipeline able to perform fully automated grading of retinal inflammation based on a standardised, clinically-validated grading scheme, namely the Tugal-Tuktun scale. The application of such scale has so far been hindered by the the amount of time required to (manually) apply it in clinical settings. Our dataset includes 3,205 fluorescein angiography images, from 148 patients and 242 eyes, from the uveitis department of Jules Gonin Eye Hospital. The data were automatically extracted from a medical device (Heidelberg Spectralis), in hospital settings. Images were graded by a medical expert. We focused specifically on one type of inflammation, namely retinal vasculitis. Our pipeline comprises both learning-based models (Pasa model with F1 score = 0.81, AUC = 0.86), and a computer vision intensity-based approach to serve as a baseline (F1 score = 0.57, AUC = 0.66). A recall of up to 0.833, computed on an independent test set, is comparable to the scores obtained by available state-of-the-art approaches. Here we present the first fully automated pipeline for the grading of retinal vasculitis from raw medical images that is applicable to a real-world clinical data.