Related Experiment Videos
MultiRetNet: A Lightweight Explainable AI Approach to Diabetic Retinopathy Grading and DME Detection Using Fundus-OCT
Saad Islam1, Ravinesh C Deo1, U Rajendra Acharya1
1Artificial Intelligence Applications Laboratory, School of Mathematics, Physics and Computing, University of Southern Queensland, Springfield, QLD 4300, Australia.
Journal of Imaging
|June 25, 2026
Summary
A new deep learning model, MultiRetNet, simultaneously screens for diabetic retinopathy (DR) and diabetic macular edema (DME) using fused eye images. This multimodal approach significantly improves DME detection sensitivity for comprehensive diabetic eye care.
Area of Science:
- Ophthalmology
- Artificial Intelligence
- Medical Imaging
Background:
- Diabetic retinopathy (DR) and diabetic macular edema (DME) are leading causes of preventable blindness.
- Current automated screening systems often analyze DR and DME in isolation, using single imaging modalities.
- This limitation hinders comprehensive and efficient screening for diabetic eye disease.
Purpose of the Study:
- To develop and evaluate a deep learning model for simultaneous DR severity grading and DME detection.
- To investigate the efficacy of fusing color fundus and optical coherence tomography (OCT) images for improved diagnostic accuracy.
- To introduce a novel multimodal architecture, MultiRetNet, for integrated diabetic eye screening.
Main Methods:
- A deep learning model (MultiRetNet) was designed using two parallel EfficientNet-B0 backbones for feature extraction from paired fundus and OCT images.
- Feature-level fusion concatenated modality-specific features into a joint representation for multi-task learning.
- The model was trained and validated on a private dataset of 425 paired eye images, with performance assessed using accuracy, AUC, sensitivity, and specificity.
Main Results:
- The fusion model achieved 82.4% accuracy for DR grading and 97.6% accuracy for DME detection on the test set.
- Compared to single-modality baselines, the fusion model demonstrated a 43% relative improvement in DME sensitivity (detecting 10/12 DME-positive eyes vs. 7/12).
- Cross-validation results corroborated these findings, with the fusion model reaching 87.1% DR accuracy and 99.1% DME accuracy.
Conclusions:
- Multimodal fusion of fundus and OCT images significantly enhances the performance of automated diabetic eye screening.
- The proposed MultiRetNet architecture offers a lightweight and effective solution for simultaneously grading DR and detecting DME.
- This approach represents a promising advancement in comprehensive diabetic eye screening, offering improved diagnostic capabilities over single-modality systems.
