Advanced Detection of Retinal Diseases Via Novel Hybrid Deep Learning Approach

Abstract

Diabetic drusen, choroidal neovascularization (CNV), and macular edema (DME) are some retinal diseases that can cause severe blindness and vision loss. Early diagnosis of retinal diseases is vital to prevent this irreversible damage to the eye. The problem statement of this study can be given as presenting new deep learning based results for detecting these retinal diseases. For this purpose, OCT dataset was used to detect CNV, DME and Drusen patients. This data set, which is frequently used in the literature, consists of CNV, DME, Drusen and Normal retina images. RestNet50, InceptionV3, InceptionResnetV2, MobileNet, DenseNet-201, Xception, EfficentNetB4, EfficentNetB7 and EfficentNetV2S models of the CNN architecture were applied to the data set and the performance results of these models were obtained. Then, in order to increase the classification performance of each of these models, hyperparameter tuning was performed by reducing the learning rate by half in each epoch. Later, a hybrid version of the EfficientNetV2S and Xception convolutional neural network models, the most successful of these hyperparameter-tuned models, was developed. The performance analysis results of our proposed hybrid deep learning model are given by comparing them with traditional deep learning models in the literature. These comparison results show that the classification success of the proposed model is higher than the success of traditional deep learning models in the literature. Thus, the proposed hybrid model can shorten the clinical diagnosis time. In addition, the costs of healthcare services can be reduced by intervening in treatable diseases earlier, instead of more costly interventions in the advanced stages of the disease.