Diagnosing and Classification of Diabetic Retinopathy via Convolutional Neural Network

In this study, we propose a novel approach to diagnosing diabetic retinopathy (DR) images. DR is one of the principal reasons for blindness worldwide that the retina's blood vessels distortion causes it. Approximately ۸۰ percent of people who suffer from diabetes endure some disease stage in almost ten years. Early diagnosis of the DR is possible through regular examinations, retinal imaging, and examination of vascular variations by an expert. Due to these time-consuming and expensive stages, it becomes crucial to classify the stages and severity of DR for the recommendation of required treatments by implementing an automated diagnostic system introduced in this article. In this research, a new method with the ability to be implemented simply excluding the demand for manual features extraction, which reduces computational time and load. The proposed method improves the evaluation parameters and is more straightforward and speedy. The recommended method is based on convolutional neural networks and comprises convolution layers, followed by fully connected layers. The images' dimensions are initially reduced in the presented algorithm, and proper preprocessing is performed to enhance images contrast to feed into the deep learning network. The model consists of ۱۸ convolutional and pooling layers and specific dense layers to extract the features and structures intending to diagnose the disease as well as determine its severity. To demonstrate the severity of the DR, it is classified into four categories. The proposed method has achieved ۹۲% and ۸۳% accuracy in the publicly available retinal image database, Kaggle, in DR diagnosing and classification.

کلیدواژه ها:

Diabetic Retinopathy ، Convolutional Neural Network ، Image Classification ، Retinal Image Processing