Deployment Aligned Hybrid Attention For Diabetic Retinopathy Screening With Mdrnet
DOI:
https://doi.org/10.29284/yyhkzs56Keywords:
Diabetic retinopathy, retinal disease screening, hybrid attention networks, class-imbalance learning.Abstract
Early identification of diabetic retinopathy (DR) is critical for preventing avoidable vision loss, particularly in primary-care settings where screening infrastructure is constrained by variable image quality, limited hardware resources and workforce pressure. This work introduces MDRNet, a deployment-aligned deep learning framework designed to provide clinically coherent DR grading under these real-world conditions. The approach integrates F-CAN perceptual harmonisation, a lightweight EfficientNet-B0 encoder, multi-scale hybrid attention, and class-distribution regularisation, yielding a model tailored for commodity CPU hardware without reliance on transformer-scale capacity. Evaluation on a patient-level EyePACS split that preserves natural epidemiological imbalance demonstrates robust multiclass performance, achieving 91.3% accuracy, 0.86 macro-F1, 0.90 quadratic weighted kappa, and 0.95 macro-AUC across the five-grade clinical scale. Referral-tier mapping, which collapses severe and proliferative disease into a clinically actionable category, achieved 0.92 sensitivity and 0.85 specificity, indicating suitability as a triage-supportive system. Calibration metrics (Brier score = 0.082; ECE = 0.034) further suggest reliability of probabilistic outputs for threshold-based decision support. The findings demonstrate that prioritising calibration, ordinal stability and inference feasibility can yield competitive performance without specialised accelerators, supporting the potential for MDRNet to contribute meaningfully to scalable DR screening pathways subject to prospective validation.
References
1. C. Jiang, X. He, Y. Zhu, and L. Tao, ‘Chinese burdens and trends of diabetic retinopathy 1990–2021 and 15 years forecast: results from the Global Burden of Disease Study 2021’, Front. Endocrinol., vol. 16, p. 1573581, May 2025, doi: 10.3389/fendo.2025.1573581.
2. Z. Li et al., ‘An Automated Grading System for Detection of Vision-Threatening Referable Diabetic Retinopathy on the Basis of Color Fundus Photographs’, Diabetes Care, vol. 41, no. 12, pp. 2509–2516, Dec. 2018, doi: 10.2337/dc18-0147.
3. N.-A. Fasoula et al., ‘Clinical and Translational Imaging and Sensing of Diabetic Microangiopathy: A Narrative Review’, JCDD, vol. 10, no. 9, p. 383, Sept. 2023, doi: 10.3390/jcdd10090383.
4. S. Divya, L. Padma Suresh, and A. John, ‘Hybrid Optimization Algorithm-Based Generative Adversarial Network for Change Detection Using Pre-Operative and Post-Operative MRI’, Int. J. Patt. Recogn. Artif. Intell., vol. 36, no. 07, p. 2251007, June 2022, doi: 10.1142/S0218001422510077.
5. S. Divya, L. Padma Suresh, and A. John, ‘Enhanced deep-joint segmentation with deep learning networks of glioma tumor for multi-grade classification using MR images’, Pattern Anal Applic, vol. 25, no. 4, pp. 891–911, Nov. 2022, doi: 10.1007/s10044-022-01064-5.
6. K. B. Nielsen, M. L. Lautrup, J. K. H. Andersen, T. R. Savarimuthu, and J. Grauslund, ‘Deep Learning–Based Algorithms in Screening of Diabetic Retinopathy: A Systematic Review of Diagnostic Performance’, Ophthalmology Retina, vol. 3, no. 4, pp. 294–304, Apr. 2019, doi: 10.1016/j.oret.2018.10.014.
7. T. Zhu, K. Li, P. Herrero, and P. Georgiou, ‘Deep Learning for Diabetes: A Systematic Review’, IEEE J. Biomed. Health Inform., vol. 25, no. 7, pp. 2744–2757, July 2021, doi: 10.1109/JBHI.2020.3040225.
8. J. Y. Zhang et al., ‘How to Best Protect People With Diabetes From the Impact of SARS-CoV-2: Report of the International COVID-19 and Diabetes Summit’, J Diabetes Sci Technol, vol. 15, no. 2, pp. 478–514, Mar. 2021, doi: 10.1177/1932296820978399.
9. M. Narasimharao, B. Swain, R. Priyadarshi, P. P. Nayak, and S. Bhuyan, ‘A Survey of Machine Learning Techniques for Diabetes Prediction: Current Trends and Future Directions’, Arch Computat Methods Eng, Oct. 2025, doi: 10.1007/s11831-025-10450-1.
10. J. Hatherley, ‘Data over dialogue: Why artificial intelligence is unlikely to humanise medicine’, 2025, doi: 10.48550/ARXIV.2504.07763.
11. U. H. Okwudili, O. O. Ukachukwu, V. C. Chijindu, M. O. Ezea, and B. Ishaq, ‘An improved performance model for artificial intelligence-based diabetes prediction’, Journal of Electrical Systems and Inf Technol, vol. 12, no. 1, p. 25, June 2025, doi: 10.1186/s43067-025-00224-x.
12. T. Zhu, K. Li, P. Herrero, and P. Georgiou, ‘Deep Learning for Diabetes: A Systematic Review’, IEEE J. Biomed. Health Inform., vol. 25, no. 7, pp. 2744–2757, July 2021, doi: 10.1109/JBHI.2020.3040225.
13. W. Zhang et al., ‘Automated identification and grading system of diabetic retinopathy using deep neural networks’, Knowledge-Based Systems, vol. 175, pp. 12–25, July 2019, doi: 10.1016/j.knosys.2019.03.016.
14. G. Rajarajeshwari and G. C. Selvi, ‘Application of Artificial Intelligence for Classification, Segmentation, Early Detection, Early Diagnosis, and Grading of Diabetic Retinopathy From Fundus Retinal Images: A Comprehensive Review’, IEEE Access, vol. 12, pp. 172499–172536, 2024, doi: 10.1109/ACCESS.2024.3494840.
15. P. Xia et al., ‘Generalizing to Unseen Domains in Diabetic Retinopathy with Disentangled Representations’, in Medical Image Computing and Computer Assisted Intervention – MICCAI 2024, vol. 15010, M. G. Linguraru, Q. Dou, A. Feragen, S. Giannarou, B. Glocker, K. Lekadir, and J. A. Schnabel, Eds, in Lecture Notes in Computer Science, vol. 15010. , Cham: Springer Nature Switzerland, 2024, pp. 427–437. doi: 10.1007/978-3-031-72117-5_40.
16. S. Chokuwa and M. H. Khan, ‘Divergent Domains, Convergent Grading: Enhancing Generalization in Diabetic Retinopathy Grading’, in 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), Tucson, AZ, USA: IEEE, Feb. 2025, pp. 3667–3677. doi: 10.1109/WACV61041.2025.00361.
17. M. J. Reyes-Medina, V. Cantón-Habas, F. J. Sánchez-Jiménez, E. Molina-Hurtado, and M. D. P. Carrera-González, ‘New approach to optimize therapy in type 2 diabetes mellitus: the importance of subclassification’, Front. Endocrinol., vol. 16, p. 1710511, Nov. 2025, doi: 10.3389/fendo.2025.1710511.
18. D. Yan, X. Li, Y. Wang, and Z. Cai, ‘Optimized prediction of diabetes complications using ensemble learning with Bayesian optimization: a cost-efficient laboratory-based approach’, Front. Endocrinol., vol. 16, p. 1593068, June 2025, doi: 10.3389/fendo.2025.1593068.
19. Z. Zhang, H. Zhao, L. Dong, L. Luo, and H. Wang, ‘A Study on the Interpretability of Diabetic Retinopathy Diagnostic Models’, Bioengineering, vol. 12, no. 11, p. 1231, Nov. 2025, doi: 10.3390/bioengineering12111231.
20. D. Kothadiya, A. Rehman, S. Abbas, F. S. Alamri, and T. Saba, ‘Attention-based deep learning framework to recognize diabetes disease from cellular retinal images’, Biochem. Cell Biol., vol. 101, no. 6, pp. 550–561, Dec. 2023, doi: 10.1139/bcb-2023-0151.
21. X. Yu, Y. Wang, Z. Liu, and E. Jung, ‘Technological functionality and system architecture of mobile health interventions for diabetes management: a systematic review and meta-analysis of randomized controlled trials’, Front. Public Health, vol. 13, p. 1549568, Feb. 2025, doi: 10.3389/fpubh.2025.1549568.
22. R. P. Homme et al., ‘Remodeling of Retinal Architecture in Diabetic Retinopathy: Disruption of Ocular Physiology and Visual Functions by Inflammatory Gene Products and Pyroptosis’, Front. Physiol., vol. 9, p. 1268, Sept. 2018, doi: 10.3389/fphys.2018.01268.
Downloads
Published
Issue
Section
License
This work is licensed under a Creative Commons Attribution 4.0 International License.
This work is licensed under a Creative Commons Attribution 4.0 International License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
