1Professor, Department of Computer Science and Engineering, Vel Tech Rangarajan Dr.Sagunthala R&D Institute of Science and Technology, Avadi, Chennai, India
Email: drsadishsendilm@veltech.edu.in
2Professor, Department of Artificial Intelligence and Data Science, Vel Tech High Tech Dr.Rangarajan Dr.Sakunthala Engineering College, Avadi, Chennai, India
Email: ksankar@velhightech.com
3Professor and Dean R & D, Department of CSD, Siddhartha Institute of Engineering and Technology (Autonomous), Ibrahimpatnam, Hyderabad, India
Email: drsaleemprincipal@gmail.com
Diabetic retinopathy (DR) remains a leading cause of preventable vision loss, yet real-world screening programs still struggle with delayed presentation, grader variability, and limited access to retinal specialists. This paper proposes an AI- and ML-powered early detection pipeline that couples deep representation learning with clinically aligned decision support to improve referable DR identification and grading from color fundus photographs. The approach integrates standardized image quality control, lesion-aware preprocessing, and a lightweight deep convolutional backbone augmented with attention to emphasize microaneurysms, hemorrhages, and exudates across multiple scales. To enhance robustness across devices and populations, training is formulated as a multi-domain problem with balanced sampling, mixup-style augmentation, and calibration-aware loss functions that explicitly penalize overconfident errors. Model outputs are mapped to clinically actionable states (no DR, mild, moderate, severe/proliferative, and referable DR) and accompanied by uncertainty estimates and saliency-based explanations to support clinician trust and triage. Evaluation is designed around decision-making needs, reporting sensitivity at fixed specificity for referable DR, quadratic weighted kappa for severity grading, and failure-to-refer risk under realistic image-quality constraints. The intended deployment is a "human-in-the-loop" workflow where the model prioritizes high-risk cases, flags ungradable images, and generates structured reports that can be reviewed by graders or ophthalmologists. By unifying accurate detection, interpretability, and workflow integration, the proposed framework aims to reduce missed disease, accelerate referrals, and strengthen clinical decision-making in resource-constrained screening settings. A prospective validation plan measures turnaround time, referral adherence, and cost per detected case, translating benchmark accuracy into measurable population benefit across diverse primary-care settings.
Keywords: Diabetic retinopathy, Deep learning, Fundus imaging, Clinical decision support, Explainable AI, Screening triage.
How to cite this article: Murugaraj SS, Sankar K, Saleem PAA. AI and ML powered early detection of diabetic retinopathy: a deep learning approach for improved clinical decision-making. Int J Drug Deliv Technol. 2026;16(8s): 77-82; DOI: 10.25258/ijddt.16.8s.15
Source of support: None.
Conflict of interest: None