Artificial Intelligence as a Surrogate for Electroretinography in Cone Dystrophy Diagnosis: A Feasibility Framework

This OSF project presents a feasibility study evaluating whether artificial intelligence-derived features from multimodal retinal imaging can estimate cone-system dysfunction in cone dystrophy and cone-rod dystrophy. The primary outcome was 30-Hz flicker full-field electroretinography amplitude. Retinal images from OCT, fundus autofluorescence, and colour fundus photography were analysed using an EfficientNetB0-based deep-learning workflow. The curated dataset included 63 ERG-linked cone dystrophy / cone-rod dystrophy cases and 665 retinal images across training and test datasets. Modality-specific CNN classifiers were first trained to distinguish disease from normal retinal images. Image features were then extracted and used for ridge-regression prediction of 30-Hz ERG amplitude. The strongest result was achieved by multimodal OCT + FAF + fundus-photo fusion, which showed a strong association with measured ERG amplitude. The multimodal model achieved R² = 0.695, Pearson r = 0.842, Spearman r = 0.844, and mean absolute error = 11.68 µV. These findings support the feasibility hypothesis that retinal images contain AI-extractable information related to cone-system electrophysiological dysfunction. This project is exploratory and does not claim that AI can replace ERG. The current model is a proof-of-concept framework requiring larger prospective validation using standardized multimodal imaging and ISCEV-compliant ERG. OSF DOI: 10.17605/OSF.IO/NQSKD