Image Processing, Textural Feature Extraction and Transfer Learning based detection of Diabetic Retinopathy

Diabetic Retinopathy (DR) is one of the most common causes of blindness in adults. The need for automating the detection of DR arises from the deficiency of ophthalmologists in certain regions where screening is done, and this paper is aimed at mitigating this bottleneck. Images from publicly available datasets STARE, HRF, and MESSIDOR along with a novel dataset of images obtained from the Retina Institute of Karnataka are used for training the models. This paper proposes two methods to automate the detection. The first approach involves extracting features using retinal image processing and textural feature extraction, and uses a Decision Tree classifier to predict the presence of DR. The second approach applies transfer learning to detect DR in fundus images. The accuracies obtained by the two approaches are 94.4% and 88.8% respectively, which are competent to current automation methods. A comparison between these models is made. On consultation with Retina Institute of Karnataka, a web application which predicts the presence of DR that can be integrated into screening centres is made.