Retinal Fundus Image Classification Using Hybrid Deep Learning Model

Abstract

Diabetic retinopathy (DR) is a disease related with diabetes. DR causes lesions on the eye retina that can affect the human vision. DR needs to be detected at its earlier stage, otherwise in the worst situation patients suffer from DR may lead to blindness. Currently, DR is diagnosed using fundus images of the retina using manual methods by medical practitioners. Automated methods are more reliable, cost-effective, and time savers as compared to their manual counterpart. In this work, a deep learning (DL) model is proposed to automatically recognize DR images by classifying retinal fundus images into five different classes. A hybrid DL model of three different DL pretrained models AlexNet, VGGNet, and ResNet-18 is used. Then the performance of the combined proposed model is enhanced using the AdaBoost algorithm. AdaBoost algorithm is used to exploit dependency among different DL models. It can handle missing values and outliers. The APTOS and Messidor-2 dataset are used for the training of all pre-trained models and proposed model. Then performance of the proposed hybrid DL model along with three other pretrained DL models is evaluated based on performance metrics accuracy, sensitivity, specificity, and F1score. From the obtained results it is to be noted that performance metric values for the proposed hybrid model are at higher side as compared to the other pretrained DL models. Hence, the proposed method is a potential method for the DR image classification.