Automated Detection of Systemic and Ophthalmic Diseases Using RETFound in Fundus Imaging

This repository contains the implementation and research materials for a scientific-practical study employing deep learning methods to analyze retinal fundus images for automated diagnosis of systemic (hypertension, diabetes) and ophthalmic (cataract, glaucoma, diabetic retinopathy, and age-related macular degeneration) diseases. Utilizing the advanced RETFound model, this interdisciplinary project enables comprehensive detection of various systemic and ocular diseases through color fundus image interpretation.

📝Key Features

• Interdisciplinary approach: Combining medical knowledge with advanced IT solutions for intelligent diagnostics.
• Multi-disease detection: Identifies systemic and ophthalmic diseases from CFI.
• State-of-the-art technologies: Powered by the RETFound model, designed for medical image analysis.
• Diverse datasets: Utilization of both private clinical, ensuring real-world applicability of the model, and public datasets.
• Transfer learning: Fine-tuning techniques applied to enhance generalizability and diagnostic accuracy across diverse datasets.

🗂Dataset Information

Private Dataset (developed with support from the Medical University of Białystok) The dataset contains 21,410 fundus images from 3,214 patients. Based on biomarkers, images were categorized into the following disease groups:

• hypertension (n=12911)
• diabetes (n=2863)
• cataract (n=2206)
• glaucoma (n=745)

Public Datasets

• Kaggle
• REFUGE2 (Retinal Fundus Glaucoma Challenge)
• ADAM (Automatic Detection challenge on Age-related Macular degeneration)

All datasets were split into 70% training, 15% validation, and 15% test sets.

📏Results

The RETFound model demonstrated strong performance across all datasets, particularly on public benchmarks:

Dataset	AUC	Accuracy
Hypertension	0.788	0.718
Diabetes	0.813	0.739
Glaucoma	0.856	0.777
Cataract	0.945	0.864
Kaggle	0.990	0.947
REFUGE2	0.976	0.947
ADAM	0.976	0.945

📊Visualization

For each dataset, visualizations were created including confusion matrices, AUC-ROC curves, PR-AUC graphs, and distribution diagrams. You can view them here.

Fig. 1: Sample predictions on Kaggle dataset (DR, cataract, glaucoma, normal).

📍Citation

@article{zhou2023foundation,
  title={A foundation model for generalizable disease detection from retinal images},
  author={Zhou, Yukun and Chia, Mark A and Wagner, Siegfried K and Ayhan, Murat S and Williamson, Dominic J and Struyven, Robbert R and Liu, Timing and Xu, Moucheng and Lozano, Mateo G and Woodward-Court, Peter and others},
  journal={Nature},
  volume={622},
  number={7981},
  pages={156--163},
  year={2023},
  publisher={Nature Publishing Group UK London}
}