An Improved Diabetic Retinopathy Image Classification by Using Deep Learning Models

Abstract

Diabetic Retinopathy (DR) is a kind of diabetes complication that damages the light-sensitive tissues of the blood vessels at the back of the eyes. Early detection of such problems along with controlling diabetes can prevent severe damages from the disease. Detection of DR is time-consuming, and manual detection is error-prone. Hence, in the majority of the cases, it is detected at a severe stage making it difficult to treat properly. To handle this problem, this paper presents a deep learning model consisting of AlexNet, VGGNet, and modified VGGNet, and ResNet, to detect DR from images. A detailed comparison among the adopted models and the state-of-the-art reveals that the modified VGGNet outperforms other applied models with 87.69% accuracy, 87.93% precision, and 87.81% recall. The model accuracy increases to 95.77% after performing hyperparameter tuning. The experimental results are promising and make the model a suitable candidate for automated DR detection from fundus images.