OR33-04 A Simple Mobile Artificial Intelligence Retina Tracker (SMART) Powered by Efficient Deep Learning Models for Diagnosis and Prognosis of Diabetic Retinopathy

Abstract Disclosure: R. Elangovan: None. K. Elangovan: None. J.R. Sethuraj: None. E. Krishnan: None. U. Qureshi: None. F. Rahman: None. R.N. Vaghela: None. Z. Baloch: None. T. Patel: None. S.H. Chaudhary: None. R. Muralidhar: None. C. Robert: None. S. Biswas: None. G. Rajamani: None. N. Vora: None. K.J. Patel: None. M. Amro Alrouh: None. A.S. Agrawal: None. S. Malapati: None. T. Dontulwar: None. K. Noor: None. F. Madhumithaa Jagannathan: None. P. Uvaraj: None. M. Kaur: None. S. Siddiq: None. M. Zeeshan: None. V. Makarla: None. S. Tamma: None. S.S. Bhurchandi: None. S.K. Bhurchandi: None. Introduction: Diabetic Retinopathy (DR), a leading cause of preventable blindness, affects millions globally as the incidence and prevalence of diabetes continue to rise worldwide. The eye’s unique accessibility for imaging, combined with artificial intelligence (AI), offers transformative opportunities for DR diagnosis and management. By leveraging oculomics—analyzing retinal images to assess systemic health—AI provides a revolutionary approach to combat vision loss while addressing broader healthcare challenges, especially in underserved regions. Aim: To develop a computationally efficient, highly accurate AI-powered model integrated into Simple Mobile AI Retina Tracker (SMART)—a universally accessible application—for precise detection, staging, and monitoring of diabetic retinopathy progression, while ensuring privacy and real-world usability. Methodology: We employed pretrained models from cutting-edge deep learning architectures (EfficientNets, ResNets, and Vision Transformers) trained on anonymized fundoscopic images from diverse diabetic populations. The trained models underwent rigorous internal validation to ensure reliability and were externally validated on multiple datasets (APTOS, JSIEC, IDRiD, MESSIDOR), confirming robust generalizability across populations. Additionally, we evaluated the models’ ability to differentiate DR from 39 ocular conditions such as hypertensive retinopathy and retinal vein occlusion. A dedicated online platform was developed to enable users to test the models with their own data locally. Multi-national healthcare professionals validated the platform’s reliability and usability, underscoring its global applicability. Results: EfficientNetB0 achieved over 99% accuracy with an image processing speed of <1 second per image across diverse datasets. It demonstrated superior computational efficiency (one-third runtime compared to ResNet18) while maintaining robust generalization across all tested datasets (AUROC >0.99). External evaluations confirmed its real-world feasibility, making it ideal for scalable deployment in low-resource settings. Conclusions: This study establishes a paradigm shift in AI-driven medicine, demonstrating that computationally efficient models like EfficientNetB0 can achieve state-of-the-art accuracy while being resource-efficient. Its integration into the SMART application ensures universal accessibility, privacy preservation, and scalability for real-world applications. Applications: By democratizing eyecare through free mobile technology, this innovation has the potential to screen billions globally, reducing vision loss from DR while transforming healthcare delivery systems. The AI-powered application is scalable as a universal tool for ocular and systemic conditions, ensuring global impact in both diagnosis and disease monitoring. Presentation: Monday, July 14, 2025