DefinePK

Abstract

Diabetic Retinopathy (DR) remains of the leading causes of visual loss worldwide; therefore early detection is primary reason for prevention of blindness. Using deep CNNs, we provide an algorithmic framework for DR screening using retinal fundus images.  By means of transfer learning a pre-trained VGG16 model trained on ImageNet was fine-tuned to identify the degree of severity of DR. Regularization methods including early stopping and dropout were used to improve generalization of model. The framework was evaluated on a wide-spectrum public dataset of DR fundus images, obtaining more than 90% accuracy on classifying healthy and DR retinas. Results indicated high sensitivity, especially in early stages of DR for early diagnosis and treatment. Comparison with recent developments in medical imaging (2024–2025) show the effectiveness of our proposed approach. Given its excellent performance and ease of implementation, this system appears to be a promising efficient tool, especially in low-resource or remote clinical settings. The review ends with acknowledging the promise of DL and potential to improve the diagnosis of DR and possible ides of future, e.g. including multi-modal images and getting alignment with clinical workflow.