Multilevel Classification Model for Diabetic Retinopathy

Diabetic retinopathy is the prominent cause of blindness in the working population of the developed world [11]. It is estimated to affect over 314 million people in the next two years. Exudates, micro aneurysms, augmented blood vessels, fluid drip, cotton wool spot, hemorrhages are preliminary signs of diabetic retinopathy. Early detection of these signs will help chances to initiate precautionary measures and thereby avoid permanent damage to the eye sight. In this paper we present systematic approach for feature selection and a machine learning [12] model to classify DR images into five category of varying stage including normal. Dataset from Kaggle is employed for experimentation and extracted around 11 features and for each data 25 features were extracted. 11 among 25 features are selected based on examining low variance. We experimented with various classification models such as Logistic regression model, random forest classifier, Gaussian NB classifier, KNN, Decision tree classifier, Gradient boosting classifier. Results in each phase are validated by 10 fold cross validation method. We compare our results with recently reported literatures. It is found that proposed model presents out performing results with sensitivity equal to 84% for Decision Tree classifier, specificity equal to 82% for Gaussian NB, and accuracy equal to 88% for KNN. The proposed model can be effectively used to aid diagnostic decisions of stage of DR.