Eye2Gene

Abstract

Purpose: 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. The prescription and interpretation of genetic results requires phenotype–genotype recognition that only IRD experts can provide hence this has motivated AI approaches that are able to predict the probable IRD causative gene from the retinal scans of suspected IRD patients. However, these AI approaches are currently “black boxes” that do not offer any clinical interpretability nor do they provide fine‐grained phenotypic information which is essential for prognosis. We therefore sought to develop an AI algorithm capable of automatically identifying and quantifying IRD‐specific features in retinal scans. Methods: In order to build a training dataset for the AI algorithm, a grading protocol was drafted defining retinal features important in the identification of the 36 most common IRD genes. Optical Coherence Tomography (OCT) and Fundus Autofluorescence (FAF) scans were manually segmented by four graders over three rounds of grading including adjudication, feedback and protocol clarification where required. This iterative process was followed to assure good agreement, assessed using the dice score metric, between features. Features that were too difficult or laborious to annotate were converted to labels. Using the manually segmented data, an AI algorithm known as a U‐net, was trained to automatically segment 15 features. The number, size and brightness of the automatically identified features was quantified and compared across the 36 gene classes. Results: A total of 3527 scan‐features were manually annotated across scans of 36 genes. The inter‐grader dice scores ranged from 0.30 to 0.91 with an average 0.54. Features with the best agreement were anatomical features such as whole retina on OCT (0.91) and optic disc on FAF (0.87), and yielded the best predictions by the segmentation model 0.87 and 0.82, respectively. Pathological features with good inter‐grader agreement included ellipsoid zone loss (0.78) and hypo autofluorescence (0.65) and those with poorer inter‐grader agreement included hyper autofluorescence (0.30) and retinal pigment epithelium loss (0.42). The segmentation model achieved an average dice score of 0.71 across all features. Statistically significant differences were found in feature count, size and brightness between the 36 different gene classes and prediction accuracy of 58% was achieved using random forests to predict the correct causative gene from 36 genes using these features. Using a “black box” AI approach the gene prediction accuracy was 66%. Conclusions: Automated segmentation of features in IRD scans using AI is feasible and leads to interpretable prediction of disease‐associated IRD genes. However, “black box” prediction can still achieve higher accuracy at the expense of interpretability. Fur...