Diabetic Retinopathy Prediction from Retinal Fundus Images using a Conditional Self-Attention GAN

In the past years, diabetic retinopathy (DR) has become one of the standard diseases of the eyes due to the maximum increase in blood glucose levels. It is estimated that around 50% of the population below 70 years suffers from severe diabetes leading to various complications like DR. Preservation of vision in patients suffering from DR happened due to early detection and proper treatment. Hence, it is very important to identify the early signs of it and rightly assess the severity of the disease to provide appropriate treatment. This paper classifies different severity levels of DR fundus images using a deep learning (DL) algorithm to initiate their early diagnosis and treatment. This paper proposes a conditional self-attention generative adversarial network for DR prediction through retinal fundus images (CSA-GAN-DRP-RFI). Preprocessing, segmentation, feature extraction, and classification are the steps in the suggested methodology. A pre-processing stage of contrast enhancement has been carried out in a fundus image from the MESSIDOR dataset through the implementation of Normalized Gamma-Corrected Contrast-Limited Adaptive Histogram Equalization (NG-CC-LAHE). Kernelized Gravity-based Density Clustering (KGDC) is used for blood vessel and optical disc segmentation. Texture features, Shannon entropy (SE), Kapur entropy (KE), and Renyi entropy (RE) are extracted with the help of Adaptive and Concise Empirical Wavelet Transform. The CSA-GAN classifies these features into the following severity levels thereafter: Mild, Moderate, and severe Non-Proliferative Diabetic Retinopathy (MiNPDR, MoNPDR, SNPDR), Proliferative Diabetic Retinopathy (PDR), and Normal. This paper proposes a hybrid optimization technique for improving the classification accuracy by tuning the weight parameters of the CSA-GAN. The performance of the suggested technique is assessed using Sensitivity, Accuracy, precision, Recall, ROC, and F-1 Score metrics on the Python platform. The performance of the proposed CSA-GAN-DRP-RFI method provides 7.12%, 6.29%, and 3.11% higher accuracy; 1.27%, 2.09%, and 2.75% higher recall comparing it with existing techniques.