Related Experiment Video
Updated: May 5, 2026
Using Retinal Imaging to Study Dementia
Published on: November 6, 2017
Diabetic Retinopathy Assessment through Multitask Learning Approach on Heterogeneous Fundus Image Datasets
Hongkang Wu1, Kai Jin1, Yiyang Jing2
1Zhejiang Provincial Key Laboratory of Ophthalmology, Zhejiang Provincial Clinical Research Center for Eye Diseases, Zhejiang Provincial Engineering Institute on Eye Diseases, Eye Center of Second Affiliated Hospital, School of Medicine, Zhejiang University, Hangzhou, China.
Objective:
To develop and validate an artificial intelligence (AI)-based system, Diabetic Retinopathy Analysis Model Assistant (DRAMA), for diagnosing diabetic retinopathy (DR) across multisource heterogeneous datasets and aimed at improving the diagnostic accuracy and efficiency.
Design:
This was a cross-sectional study conducted at Zhejiang University Eye Hospital and approved by the ethics committee.
Subjects:
The study included 1500 retinal images from 957 participants aged 18 to 83 years. The dataset was divided into 3 subdatasets: color fundus photography, ultra-widefield imaging, and portable fundus camera. Images were annotated by 3 experienced ophthalmologists.
Methods:
The AI system was built using EfficientNet-B2, pretrained on the ImageNet dataset. It performed 11 multilabel tasks, including image type identification, quality assessment, lesion detection, and diabetic macular edema (DME) detection. The model used LabelSmoothingCrossEntropy and AdamP optimizer to enhance robustness and convergence. The system's performance was evaluated using metrics such as accuracy, sensitivity, specificity, and area under the curve (AUC). External validation was conducted using datasets from different clinical centers.
Main Outcome Measures:
The primary outcomes measured were the accuracy, sensitivity, specificity, and AUC of the AI system in diagnosing DR.
Results:
After excluding 218 poor-quality images, DRAMA demonstrated high diagnostic accuracy, with EfficientNet-B2 achieving 87.02% accuracy in quality assessment and 91.60% accuracy in lesion detection. Area under the curves were >0.95 for most tasks, with 0.93 for grading and DME detection. External validation showed slightly lower accuracy in some tasks but outperformed in identifying hemorrhages and DME. Diabetic Retinopathy Analysis Model Assistant diagnosed the entire test set in 86 ms, significantly faster than the 90 to 100 minutes required by humans.
Conclusions:
Diabetic Retinopathy Analysis Model Assistant, an AI-based multitask model, showed high potential for clinical integration, significantly improving the diagnostic efficiency and accuracy, particularly in resource-limited settings.
Financial Disclosures:
The author(s) have no proprietary or commercial interest in any materials discussed in this article.
