Deep Learning for Early Detection of Diabetic Retinopathy from Low-Cost Retinal Scans: A Multi-Ethnic Cohort Study

Dr. Kenji Tanaka; Prof. Anja Müller; Dr. Aisha Khan

Abstract

Background: Diabetic retinopathy (DR) remains a leading cause of blindness globally, particularly in underserved populations. Early detection and intervention are crucial for preventing vision loss. Traditional screening methods can be resource-intensive and require specialized expertise. Artificial intelligence (AI) offers a promising avenue for scalable and accessible DR screening. This study investigates the efficacy of a novel deep learning model in accurately detecting DR from low-cost, non-mydriatic retinal images across diverse ethnic groups. The aim is to enhance early diagnosis and facilitate timely referral in resource-limited settings. Methods: A retrospective cohort study analyzed 50,000 retinal images from patients with and without DR, collected from four distinct geographic regions (Japan, Germany, Pakistan, Kenya). A convolutional neural network (CNN) was trained on a subset of the data and validated against a held-out set, with ophthalmologist annotations serving as the gold standard. Sensitivity, specificity, and AUC were primary outcome measures, along with subgroup analysis by ethnicity. Results: The deep learning model achieved a sensitivity of 94.7% (95% CI: 93.2-96.1) and a specificity of 91.2% (95% CI: 89.8-92.5) for detecting referable DR, with an AUC of 0.97. Performance remained robust across all ethnic subgroups, showing minimal variance (p > 0.05). Conclusion: Our deep learning system demonstrates high accuracy and generalizability for early DR detection from low-cost retinal scans. This technology has the potential to significantly improve access to screening and reduce preventable blindness in diverse populations.

Keywords: diabetic retinopathy, deep learning, retinal imaging, early detection, telemedicine

The global burden of diabetic retinopathy (DR) is substantial, especially in low- and middle-income countries where access to specialized ophthalmological care is often limited. Automated, AI-driven screening solutions can bridge this gap by providing rapid, cost-effective, and accurate assessments, thereby enabling earlier intervention and preventing irreversible vision loss. This study directly addresses the critical need for scalable diagnostic tools in underserved communities.

The aim of the study.

To develop and validate a robust deep learning model for the accurate and ethnically generalized detection of diabetic retinopathy from low-cost retinal images, exploring its potential for widespread implementation in primary care and remote settings.

Material and methods.

This retrospective, multicenter study utilized a dataset of 50,000 retinal fundus images (25,000 DR, 25,000 non-DR) from patients across four distinct cohorts in Japan, Germany, Pakistan, and Kenya. Images were captured using various non-mydriatic cameras. A pre-trained EfficientNetB7 convolutional neural network architecture was fine-tuned using a 70/15/15 train/validation/test split. Ophthalmologist-validated ground truth labels for referable DR were established by at least two independent certified specialists. Performance metrics included sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and the Area Under the Receiver Operating Characteristic curve (AUC). Statistical analysis, including chi-square tests and ANOVA, was performed to assess model performance across different ethnic and demographic subgroups.

The deep learning model demonstrated exceptional discriminatory power. For referable DR, the overall sensitivity was 94.7% (95% CI: 93.2-96.1), specificity was 91.2% (95% CI: 89.8-92.5), PPV was 89.6%, and NPV was 95.5%. The AUC achieved an impressive 0.97. Subgroup analysis revealed consistent performance across all four ethnic cohorts: Japanese (AUC 0.96), German (AUC 0.97), Pakistani (AUC 0.96), and Kenyan (AUC 0.95), with no statistically significant differences in performance metrics (p > 0.05). The model misclassified 3.5% of mild DR cases as healthy, but only 0.8% of severe DR cases. Processing time per image averaged 0.2 seconds on standard GPU hardware.

This study successfully developed and validated a high-performing deep learning model for the early detection of diabetic retinopathy, demonstrating remarkable accuracy and robustness across diverse populations. Its ability to process low-cost retinal images efficiently positions it as a transformative tool for improving DR screening programs globally, particularly in areas with limited access to specialists. Future work will focus on prospective clinical validation and integration into existing healthcare workflows.

1. Abadi M, et al. TensorFlow: Large-Scale Machine Learning on Heterogeneous Distributed Systems. arXiv preprint arXiv:1603.04467. 2016.

2. Gulshan V, et al. Development and Validation of a Deep Learning Algorithm for Detection of Diabetic Retinopathy in Retinal Fundus Photographs. JAMA. 2016;316(22):2402-2410.

3. Ting DSW, et al. Development and Validation of a Deep Learning System for Diabetic Retinopathy and Related Eye Diseases Using Retinal Images From Multiethnic Populations. JAMA. 2017;318(22):2211-2223.

4. Simard P, et al. Best practices for convolutional neural networks applied to visual document analysis. IJCARS. 2003;1(2):77-87.

5. Joarder M, et al. The Global Burden of Diabetic Retinopathy. Diabetes Res Clin Pract. 2020;164:108151.