Features Extraction and Concatenation of Transfer Learning Models for Early Diagnosis of Diabetic Retinopathy

Diabetic retinopathy is of interest because it is the cause of blindness in the working population and because its incidence is also increasing sharply in developing countries. Early detection of this disease is essential for a good prognosis. However, the lack of ophthalmologists and the cost of diagnosis with an increasing number of patients with diabetic retinopathy pose many problems. Rapid detection of this disease can indeed be helpful for ophthalmologists. This study deals with the development of an automatic system to classify and predict the diagnosis of diabetic retinopathy through retinal images. The system adopted is based on the technique of convolutional neural networks using a concatenation for all the feature maps of extraction of three models of transfer learning techniques such as ResNet50, InceptionV3, and EfficientNetB5 to generate a detection and prediction model. Following that, a classifier was given these features in order to classify diabetic retinopathy. The dataset used in this research is the publicly available Kaggle dataset (APTOS2019) of retina images with different preprocessing steps. The results obtained are promising, where the best achieved was 97.24% accuracy and 0.07% loss. It can be noted that the developed system has good performance compared to those published in the literature.