Early detection and classification of diabetic retinopathy by transfer learning of NASNet-large and ResNet-50 convolutional neural networks

Abstract

Diabetic Retinopathy (DR) is a progressive eye disease that affects those with long-term diabetes. It can lead to irreversible blindness if not detected and treated early. Early detection is challenging as changes to the retina are initially subtle. A number of computational models have been proposed to detect DR in fundus images, including in its early stages. Here, a novel transfer learning approach is proposed using the NASNet-Large and ResNet-50 convolutional neural networks. Image pre-processing steps are tested combinatorically. Class imbalance is addressed with oversampling and data augmentation to give trustworthy performance metrics. The models give impressive detection rates using a standard dataset containing expert-labelled DR fundus images (APTOS 2019), with the best performing models giving accuracy in classifying unseen images exceeding 0.96 (F1 score 0.97) for Early-stage DR detection (no DR vs mild and moderate), and over 0.91 accuracy (F1 score 0.91) for Multi-stage classification (no DR, mild, moderate, severe, and proliferative). This work highlights the potential of combining the transfer learning of state-of-the-art deep learning models with classical image processing for effective DR detection and classification.