Abstract
We developed an algorithm, aIRDetect, to detect, segment and quantify relevant features in fundus autofluorescence (FAF) image features in inherited retinal diseases (IRDs), to enable gene-phenotype association studies and monitoring of disease progression, at scale.
Five features of interest were defined: vessels, optic disc, perimacular ring of increased signal (ring), relative hypo-autofluorescence (hypo-AF) and hyper-autofluorescence (hyper-AF). These were manually annotated by six graders based on an agreed grading protocol to produce segmentation masks to train the AI algorithm.
Two nnUnet deep-learning models were trained on a manually segmented dataset consisting four features in 506 FAF images from 424 MEH patients and another model on segmented vessels in 72 FAF images from 52 RLH patients, validated on 48 FAF images from 40 patients at MEH, and 23 FAF images from 22 patients at RLH. Six graders manually segmented 3947 masks in total.
Model-grader dice scores for disc, hypo-AF, hyper-AF, ring and vessels were respectively 0.86, 0.72, 0.69, 0.68 and 0.65, with intergrader dice scores of 0.82, 0.75, 0.72, 0.80, 0.95, respectively.
