Related Experiment Video
Updated: Aug 5, 2026
09:17
Using Retinal Imaging to Study Dementia
Published on: November 6, 2017
Classification of Inherited Retinal Diseases Using Artificial Intelligence Models for Fundus Autofluorescence and
Han Trinh1, Ibrahim Muhammed2, Jason Charng1,3
1Department of Optometry and Vision Sciences, School of Health and Clinical Sciences, The University of Western Australia, Crawley, Western Australia, Australia, uwa.edu.au.
Journal of Ophthalmology
|August 2, 2026
Summary
Artificial intelligence models, including RETFound and ResNet, show promise for classifying inherited retinal diseases (IRDs) using fundus autofluorescence and ultra-widefield images. ResNet architectures achieved the best overall performance in this study.
Area of Science:
- Ophthalmology
- Medical Imaging
- Artificial Intelligence
Background:
- Inherited retinal diseases (IRDs) are a significant cause of vision loss in adults.
- Current artificial intelligence (AI) approaches for IRD classification face challenges with limited datasets and data labeling.
- Foundation models like RETFound offer potential solutions by leveraging large-scale pretraining.
Purpose of the Study:
- To adapt and evaluate the RETFound foundation model and convolutional neural networks (CNNs) for classifying IRDs.
- To analyze fundus autofluorescence (FAF) and pseudocolour ultra-widefield (UWF) images for IRD classification.
- To establish a robust framework for AI-driven IRD diagnosis.
Main Methods:
- Fine-tuning the RETFound model on deidentified FAF and UWF images from patients with specific IRDs and controls.
- Comparing the performance of fine-tuned RETFound against classical machine learning algorithms and other deep learning architectures (ResNet, Vision Transformer, EfficientNet_B0, ConvNeXt-Tiny).
- Utilizing datasets including Best disease, rod-cone dystrophy, Stargardt disease, and choroideremia.
Main Results:
- The fine-tuned RETFound model achieved an accuracy of 0.815, outperforming classical ML models and a pretrained Vision Transformer.
- ResNet18 and ResNet50 architectures demonstrated superior performance, achieving weighted F1 scores of 0.839 and 0.825, respectively.
- ResNet architectures showed the best overall classification accuracy for various IRDs and normal retinas.
Conclusions:
- Both ResNet architectures and the fine-tuned RETFound model exhibit high accuracy in classifying IRDs.
- These AI models hold significant potential for clinical integration in eye care settings.
- AI tools can aid in the diagnosis, triage, and management of inherited retinal diseases.