Deep Learning Methods for Segmenting and Classifying Diabetic Retinopathy

The most common form of diabetic retinopathy, or DR, which impairs vision on the retina of the eye, is diabetes. This causes blindness if not detected at the early stage because DR is not an irreversible process. This disease occurs between age group of 24-75 years. When timely treatement is given approximately we can save 90% people from their vision loss. Detection of the disease becomes difficult as there are only few symptoms so early diagnosis is required to avoid vision loss. Detection and classification of the disease is a tedious task. In the Existing systems the accuracy of the disease is calculated using only small dataset. In the proposed system we will take a large dataset where patients family medical history, diet, low and high quality images will be included. The approach also adds some extra preprocessing techniques like artifact removal, Data augmentation, and feature extraction for classification of the disease. Architectures such as ResNet, InceptionNet, or DenseNet have shown promising results in DR prediction tasks. Ensemble methods involve combining multiple models to make predictions, which often leads to improved accuracy. The bigger the dataset we take the higher the accuracy rate we achieve. The suggested methods build on existing research by utilizing deep learning algorithms and by substituting auto encoders for the dimensionality reduction PCA methodology in order to improve accuracy. Finding enough labeled data for deep learning models to train on might be difficult in the case of diabetic retinopathy. Training and deploying deep learning models require significant computational power and resources, which can be expensive and impractical for some medical facilities.