Diabetic Retinopathy Detection through Ensemble Transfer Learning Models and Web App

Abstract

Diabetic Retinopathy (DR) is a diabetic issue that disturbs the eyes. The people with diabetics will have this eye disease. Due to extreme blood sugar level, it harms the blood vessels in the retina. The blood vessel can inflates and outflows or stops the flow of blood or there will be growth of abnormal blood vessels. The DR does not show any symptoms and it will cause a severe blindness if it is not detected earlier. Detecting DR through manual process is a takes more time and requires an expert or most skilled clinician. Hence an automated detection model will solve this problem. Therefore a Deep Learning model is proposed by ensemble three transfer learning models ResNet50, VGG16 and EfficientNet-B0. The model is trained on the pre-processed retinal images and pre-processing includes cropping, applying Gaussian blur and re-sizing the images. The model will classify the different stages of DR such as mild, moderate, severe and proliferative DR. We get training accuracy of 98%, 96% and 98% for ResNet50, VGG16 and EfficientNet-B0. We obtained an 86% of testing accuracy in ensemble model. Web interface is created for detection of DR.