Loading...
Efficient deep learning models for diabetic retinopathy screening and severity assessment
Ali, Ghous Ansari, Mohammad Samar
Ali, Ghous
Ansari, Mohammad Samar
Citations
Altmetric:
Advisors
Editors
Other Contributors
Affiliation
EPub Date
Publication Date
2026-03-24
Submitted Date
Collections
Files
Accepted version
Adobe PDF, 3.46 MB
- Embargoed until 2028-03-24
Other Titles
Abstract
Diabetic retinopathy (DR) is a leading cause of preventable blindness worldwide. Early detection through systematic screening is critical, yet manual grading of fundus images by ophthalmologists is labor-intensive, subjective, and increasingly unsustainable given the rising global diabetes burden. This paper presents a dual deep learning approach to automated DR detection, addressing both binary classification (DR vs. No DR) and five-stage severity grading. We propose two complementary models: an ultra-lightweight convolutional neural network with only 11,981 parameters achieving 92% accuracy for binary classification, and a fine-tuned DenseNet121 architecture achieving 87% accuracy for multi-class severity grading. The lightweight model employs depthwise separable convolutions to minimize computational cost while maintaining high diagnostic accuracy, making it suitable for deployment in resource-constrained environments and mobile screening platforms. The DenseNet121 model leverages transfer learning with Mish activation, L2 regularization, and aggressive data augmentation to handle class imbalance on the APTOS 2019 dataset. Our results demonstrate that efficient lightweight architectures can compete with complex models for screening-level tasks, while deeper transfer learning models are essential for detailed clinical severity assessment, providing practical solutions.
Citation
Ali, G., & Ansari, M. S. (2026, January 19-21). Efficient deep learning models for diabetic retinopathy screening and severity assessment. 2026 International Conference on Al-Driven Smart Systems and Ubiquitous Computing (ICAUC) Thailand (pp. 1775-1782). https://doi.org/10.1109/ICAUC68182.2026.11441290
Publisher
IEEE
Journal
Research Unit
PubMed ID
PubMed Central ID
Type
Conference Contribution
Language
Description
“© 2026 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.”
Series/Report no.
ISSN
EISSN
ISBN
9798331558529
ISMN
Gov't Doc
Test Link
Sponsors
unfunded