Publication

Abstract : Diabetic retinopathy (DR) has become a major issue among ophthalmologists worldwide, with a majority of type 2 diabetes patients suffering from this disease. Currently, medical diagnosis is primarily performed through manual examination using an ophthalmoscope, which requires trained doctors, or other imaging devices. The scarcity of experts and the large population affected by chronic diabetes highlight the need for automated diagnostic tools—software that can provide accurate diagnostics and work effectively even with limited data. The study leveraged a pre-trained EfficientNet-B0 model, which was fine-tuned with DR-specific data to address the global issue of DR. Due to a limited dataset, advanced data augmentation techniques were implemented to enhance the model's robustness. Additionally, a customized dense layer was integrated for the precise classification of DR from levels 0 to 4. The use of Canny edge-based detection accurately segmented retinal vascular blood vessels, resulting in significantly improved classification accuracy. The methodology achieved an impressive accuracy of 87.73% across DR stages and an outstanding 95.77% precision for the No-DR class, highlighting high recall metrics for each category. This work effectively integrated state-of-the-art data augmentation and edge detection techniques into a unified system, providing a superior approach for diagnosing eye diseases such as DR that may surpass current automation standards