Diabetic Retinopathy Segmentation and Classification Using U-Net with Convolutional Neural Network

The early detection of Diabetic Retinopathy (DR) is necessary to aid doctors in treating retina patient to increase their existence rate. Due to the major imbalance dataset issue in classification, Deep Learning (DL) based approach is proposed in this research for segmentation and classification. The MESSIDOR dataset is utilized in this research for classification of Diabetic retinopathy. The min-max normalization, CLAHE and Resize is used in this experiment for data pre-processing and fed into Convolutional Neural Network (CNN) classification. The segmentation is extracted using UNet for the semantic segmentation of red lesions, semantic segmentation is achieved due to the relation of the encode and decode side feature maps. Next, Classification using CNN, they are two classifiers, normal and Mild/ moderation to classifying the object. The obtained result show that the proposed DL using CNN model achieved a better accuracy of 99.84% on MESSIDOR dataset, the existing method such as Capsule method and Moth Search Optimization model (MSO) compared to implemented model are best accuracy.