Classification of Diabetic Retinopathy (DR) using ECA Attention Mechanism Deep learning Networks

Abstract

Diabetic retinopathy (DR) is the most frequent eye condition among diabetics and a leading cause of blindness. Effective management of the disease requires regular fundus photography screening and prompt action. A computer-aided and entirely automated diagnosis of DR has attracted attention due to the increasing number of diabetic patients and the extensive screening they need. In recent years, advanced deep neural networks have been widely used in various fields. We propose to early detect the DR using the Efficient channel attention (ECA) model. For DR color medical picture severity detection, a deep convolutional neural network model called ECA-Resnet101 (ERNet), ECA-VGG19 (EVNet), and ECA-InceptionV4 (EIANet) have been constructed using a flexible one-dimensional convolution kernel size approach dependent on the feature map dimension. As a result, the Kaggle competition's DR dataset has a precision, accuracy, sensitivity, and specificity of 0.974, 0.974, 0.974, and 0.992. According to several experiments, InceptionV4, depending on the ECA, could more accurately detect disease features and define DR severity.