Classification of Inherited Retinal Diseases Using Artificial Intelligence Models for Fundus Autofluorescence and Ultrawide Retinal Images.

Background/Objectives: Inherited retinal diseases (IRDs) are a leading cause of blindness in working-age adults. Although artificial intelligence (AI) shows potential for disease classification, progress is limited by small datasets, reliance on labelled data and limited integration of multiple imaging modalities. RETFound, a foundation model pretrained on over 900,000 fundus photographs, may address these limitations. While RETFound is based on the transformer architecture, convolutional neural networks (CNNs), such as ResNet and EfficientNet_B0, have also demonstrated strong performance in classifying retinal diseases. This study adapted RETFound and CNNs for analysing fundus autofluorescence (FAF) and pseudocolour ultra-widefield (UWF) images, establishing a framework to classify IRDs. Subjects/Methods: Deidentified FAF and UWF images were obtained from controls and patients with Best disease, rod-cone dystrophy, Stargardt disease and choroideremia. Images underwent preprocessing and were used to fine-tune RETFound. Model performance was compared with classical machine learning algorithms (logistic regression, support vector machine, gradient boosting and random forest) and deep learning architectures (ResNet18, ResNet50, vision transformer, EfficientNet_B0 and ConvNeXt-Tiny) pretrained on ImageNet. Results: The fine-tuned RETFound model achieved an accuracy of 0.815, with the best performance for rod-cone dystrophy (F1 = 0.820). RETFound outperformed classical machine learning models and a vision transformer pretrained on ImageNet. ResNet18 achieved a weighted F1 of 0.839 and demonstrated the best classification performance for Stargardt disease and choroideremia (F1 0.821 and 0.728, respectively). ResNet50 achieved a weighted F1 of 0.825 and demonstrated the best classification performance for normal retinas and Best disease (F1 0.945 and 0.616, respectively). ResNet architectures achieved the best performance overall. Conclusions: ResNet architectures and fine-tuned RETFound both demonstrated strong accuracy in classifying IRD classes, demonstrating potential for clinical application and deployment in eye care settings as tools for IRD diagnosis, triage and management.