Deep Learning-Based Empirical Convolutional Methodologies for Diabetic Retinopathy Prognostication

Diabetic retinopathy (DR) is a leading cause of global blindness. The study of DR calls for a critical approach for accurate and efficient detection of DR to halt irreversible vision loss. This study proposes a new automated DR detection framework through deep learning in PyTorch using retinal fundus images. Our proposed framework uses transfer learning by fine-tuning a pre-trained convolutional neural network (CNN) architecture that was already trained on the APTOS 2019 Blindness Detection dataset. Using the proposed model, we achieved 97% accuracy and an area under the curve (AUC) of 0.95, which indicates that our model produces competitive results for detecting DR at different variations in imaging parameters. The framework has broad applicability and includes data augmentation, preprocessing, and hyperparameter tuning to improve the robustness of DR automated detection models and to improve generalization. We present a scalable approach to clinical diagnostic work, achieving high sensitivity and specificity to identify the DR process at the earliest and in an accurate manner.