Deep Learning for Early Diagnosis of Diabetic Retinopathy: a Study Using Convolutional Neural Network

This study presents a novel approach for the automated diagnosis of diabetic retinopathy using deep learning models. The proposed method includes preprocessing and preparation of retinal images, transfer learning with Inception network for retinopathy classification, training of U-Net for retinal image segmentation, multiclassification of segmented images with a custom-designed CNN, and merging of CNN models for improved performance. The APTOS 2019 Blindness Detection and DRIVE datasets were used for training and testing. The results showed that the Inception network achieved an accuracy of 87% for retinopathy classification, the U-Net model achieved a dice coefficient of 0.95 for retinal image segmentation, and the proposed CNN model achieved an accuracy of 83% for multiclassification of segmented images. The merged (ensembled) CNN models resulted in an overall accuracy of 88.7% for retinopathy classification and segmentation. The proposed method outperformed previous works in terms of accuracy and robustness. The findings demonstrate the potential of deep learning models for accurate and efficient diagnosis of diabetic retinopathy.