Diabetic Retinopathy Detection with Feature Enhancement and Deep Learning

Abstract

Diabetic Retinopathy is a medical condition which occurs in people having diabetes. Both type 1 and type 2 diabetes patients are affected by this disease. The major problem is there are no significant symptoms of this disease at its early stages for it to be identified and treated, only when it gets worse it starts to show symptoms. Primarily, there are two major types of diabetic retinopathy which are non-proliferative diabetic retinopathy and proliferative diabetic retinopathy. Non-proliferative diabetic retinopathy is the early stages of the disease which when identified can be treated. Proliferative diabetic retinopathy is the later stage of the disease which when identified is difficult to treat. This disease is both time taking and prone to errors when manually examined by doctors. Deep learning has been in existence for a while and has been efficient in analysing medical images. Convolutional Neural Networks (CNNs) are used for feature extraction from the image and difference of gaussians algorithm is used in improving feature extraction. The dataset used is from Kaggle containing around 35126 images. There are some deep learning models for this problem statement but they don't predict the disease flawlessly. Three convolutional neural networks are trained, validated and tested on the dataset. These models perform better than the existing models in predicting diabetic retinopathy on the Kaggle dataset.