RETICAD: A tool for generating dye-free fluorescein angiography and quantifying retinal disease activity

Abstract

Purpose : Fluorescein angiography (FA) visualizes functional changes in retinal vasculature that are critical for the diagnosis and management of retinal diseases (e.g., hypoperfusion, leakage from retinal microaneurysms and microvessels, and neovascularization). However, the method has two major drawbacks: (1) it requires an intravenous injection of a fluorescent dye, and (2) there are no quantitative tools for FA analysis. Hence, the use of FA has seen a steady decline with the increase in the availability of dye-free and quantitative optical coherence tomography (OCT). OCT, nonetheless, has its own limitations, such as the inability to visualize leakage and retinal periphery. The goal of this work was to address the limitations of FA by developing: (1) a dye-free equivalent of FA; and (2) quantitative measurement of FA-based disease markers.

Methods : To develop dye-free FA, we designed an AI framework that generates a sequence of FA images from a single fundus image. The framework was trained on a large dataset of FAs and corresponding fundus images from several retinal diseases (i.e., diabetic retinopathy [DR], age-related macular degeneration [AMD] and retinal vein occlusion [RVO]). To achieve automated FA quantification, we developed an image-processing algorithm that quantifies three FA-based parameters of retinal disease activity: retinal blood-flow, microvascular leakage, and tissue perfusion. This algorithm also allowed us to evaluate the similarity between real and AI-generated FAs. Collectively, these two software modules were named RETICAD.

Results : Our preliminary results suggest that AI-generated FA is highly comparable to real FA in measures of microvascular leakage, blood flow, and retinal perfusion (Fig. 1).

Conclusions : These findings highlight the feasibility of performing dye-free and quantitative FA in any setting with a fundus camera and a computer.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.

 

Fig. 1: Representative examples of AI-generated FA and quantitative FA maps.