Related Experiment Video
Updated: May 10, 2026
09:17
Using Retinal Imaging to Study Dementia
Published on: November 6, 2017
21.6K
Deep learning generalization for diabetic retinopathy staging from fundus images
Yevgeniy Men1,2, Jonathan Fhima2,3, Leo Anthony Celi4,5,6
1Andrew and Erna Viterbi Faculty of Electrical & Computer Engineering, Technion, Israel Institute of Technology, Haifa 3200003, Israel.
Physiological Measurement
|January 9, 2025
Summary
A new deep learning model, DRStageNet, accurately stages diabetic retinopathy (DR) from fundus images, overcoming generalization challenges across different datasets for improved vision loss detection.
Area of Science:
- Ophthalmology
- Medical Imaging
- Artificial Intelligence
Background:
- Diabetic retinopathy (DR) is a leading cause of vision loss, necessitating early detection.
- Current AI models for DR staging from fundus images often fail to generalize across diverse datasets due to domain shifts.
Purpose of the Study:
- To develop and validate DRStageNet, a deep learning model for robust DR staging.
- To address the generalization limitations of existing DR detection algorithms.
Main Methods:
- Developed DRStageNet using six public datasets comprising 91,984 digital fundus images (DFIs).
- Benchmarked five self-supervised vision transformers (ViTs), selecting DINOv2 for further training.
- Employed a multi-source domain (MSD) fine-tuning strategy to enhance model generalization.
Main Results:
- DINOv2 achieved a 27.4% L-Kappa improvement over other ViTs.
- MSD fine-tuning boosted performance in four out of five target domains.
- DRStageNet correctly classified 77.5% of mislabeled images, indicating high accuracy.
Conclusions:
- DRStageNet demonstrates accurate diabetic retinopathy staging, effectively tackling cross-domain generalization issues.
- The developed model and its explainability heatmaps offer a promising tool for clinical application.
