Purchase this article with an account.
NIKOLAS PONTIKOS, William Woof, Peter Krawitz, Gavin Arno, Kristina Hess, Malena Daich Varela, Bart Liefers, Thales Guimaraes, Mital Shah, Savita Madhusudhan, Susan M. Downes, Konstantinos Balaskas, Omar Abdul Rahman Mahroo, Frank G Holz, Michel Michaelides, Andrew R Webster; Eye2Gene: prediction of causal inherited retinal disease gene from multimodal imaging using AI. Invest. Ophthalmol. Vis. Sci. 2022;63(7):1161.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Inherited retinal diseases (IRDs) are single-gene disorders caused by genetic mutations in any one of over 270 genes. Identifying the causative gene through genetic testing is crucial for gene targeted treatments, recruitment to clinical trials, prognosis and family planning. However the prescription and interpretation of genetic results requires phenotype-genotype recognition that only few IRD experts can provide. Therefore we aimed to develop Eye2Gene, an AI algorithm, to predict the probable IRD causative gene from the retinal scans of suspected IRD patients.
Eye2Gene was trained and tested on retinal scans of IRD patients with a known genetic diagnosis from Moorfields Eye Hospital (MEH). Following quality control, the MEH training dataset consisted of 44,817 images from 1,907 IRD patients from MEH, covering 3 modalities: Fundus Auto-Flourescence (FAF), Infrared (IR), and Spectral-Domain Optical Coherence Tomography (SD-OCT). For each of the 3 modalities, five distinct InceptionV3 convolutional neural networks (CNNs) were trained on different subsets of the training data using 5-fold cross validation to identify up to 36 gene classes. This resulted in Eye2Gene, an ensemble of 15 CNNs. Generalisability was assessed on a held-out dataset consisting of 264 patients from MEH and an external cohort of 37 patients from University Hospital of Bonn (UHB). To benchmark Eye2Gene against human performance, a subset 50 FAF scans were evaluated by 8 ophthalmologists.
Eye2Gene yields a top-5 accuracy of 88% in the MEH held-out dataset and 83% in the external validation UHB dataset. On the human benchmarking dataset, Eye2Gene achieved a top-5 accuracy of 72% compared to 78% for ophthalmologists.
Eye2Gene is an AI algorithm capable of predicting the 36 top most common IRD genes to a top-5 accuracy of >80%. Eye2Gene achieves performance similar to a consensus of human experts on an external dataset. Eye2Gene can eventually enable democratisation of IRD expertise, currently only available in a few centres around the world.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.
A] Eye2Gene provides IRD-gene prediction for up to 36 genes given retinal scan from FAF, IR or SD-OCT imaging modalities at overall top-5 accuracy of 88% (internal test). B] Receiver-Operator Characteristic curves per imaging modality across the 36 gene classes. Percentages relate to total images per gene.
C] Eye2Gene rediction counts vs actual count D] Saliency maps
This PDF is available to Subscribers Only