New accurate deep learning model for diabetic retinopathy detection utilizing sequential pre-processing and transfer learning

Abstract

This study considers the detection of Diabetic Retinopathy (DR) using deep learning. DR affects 80% of diabetic patients worldwide and is the second leading cause of blindness. Many studies have shown that early diagnosis and treatment are critical to prevent disease progression. The contribution of this study is the development of an accurate DR detection algorithm and corresponding model, where hemorrhages were brought to a more apparent form by an efficient processing pipeline. To handle limited DR data resources and to make the visibility of bleeding in the eye more apparent, we have developed an efficient deep learning model by combining data augmentation, pre-processing, transfer learning, and adaptation of a compatible CNN. The employed dataset comprises fundus images of individuals, which are categorized into five stages of DR. For evaluation, comprehensive ablation studies are conducted on the model. Next, the developed model is evaluated against state-of-theart algorithms and demonstrates promising results in key metrics. Particularly, the model yields 96.95% accuracy and introduces a false negative rate below 1%. Efficient metrics of the study minimize the risk of missed diagnoses and reduce the likelihood of severe vision loss in diabetic patients. Therefore, our model has the potential to contribute to clinical patient care.