Creating the world’s largest dataset of segmented Inherited Retinal Disease features by bootstrapping manual annotations with AI

Abstract

Purpose : Diagnosis and analysis of Inherited Retinal Diseases (IRDs) relies on phenotype-genotype recognition, typically involving the identification of pathological and physiological features of the retina within retinal images across a variety of scanning modalities. Unfortunately manually segmenting these features is a time-consuming process, and can typically only be performed by trained specialists. We leverage deep-learning based approaches by training an AI algorithm to automatically segment a large dataset of IRD scans, producing the world’s largest database of automatically annotated IRD features.

Methods : In order to build a training dataset for the AI algorithm, a grading protocol was drafted defining retinal features important in the identification and differentiation of different IRD genes. Optical Coherence Tomography (OCT) and Fundus Autofluorescence (FAF) scans from our database of IRD patients were manually segmented by four graders over three rounds of grading. Using the manually segmented data, we trained a variety of deep-learning-based segmentation and classification models for the different image features and labels. We then applied these models to the remaining scans within our dataset to obtain a large database of automatic annotations.

Results : A total of 2283 (1023 FAF, 1260 OCT) images were manually annotated across 1215 scans (1001 FAF, 214 OCT). The inter-grader DICE scores ranged from 0.692 to 0.983 with an average 0.864. Model-grader DICE scores ranged from 0.46 to 0.95 with an average of 0.71. Models were applied to a total of 2,015,842 images (136,631 FAF, 1,879,211 OCT) across 105,162 scans (52,508 FAF, 52,703 OCT). We compared a number of different values across genes, and in other settings. For example we find that mean RPE-loss was higher in patients with variants in genes associated with Retinitis Pigmentosa phenotypes, confirming clinical findings.

Conclusions : Automated segmentation of features in IRD scans using AI is feasible, and can be used to effectively analyse large scale datasets of retinal images. We hope to extend this work in future to include additional image features e.g. retinal layers, flecks, vessels, and to other devices and imaging modalities.

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