Transfer Learning based Diabetic Retinopathy Classification

Abstract

The domain Artificial Intelligence and Machine learning is one among the budding and has become one of the important and emerging fields. In a heavily populated country like India, bringing automation in the healthcare domain is still a great challenge. Diabetic retinopathy (DR) is a common cause of blindness among working aged adults. DR is classified into 5 categories namely No DR, Mild, Moderate, Severe and Proliferative DR. Currently in many places fundus images are captured from rural areas and then with the help of trained doctors, the stage is determined but it needs automation as the above mentioned process takes time and is highly human dependent. Also as the time taken for manual detection is more, there are a lot of chances that many patients cannot be identified at the earliest which leads to increase in blindness among people. Therefore a system is needed which could automate this manual process. Transfer based models namely Densenet201, InceptionV3 and Resnet50 are implemented with the kaggle dataset. The Quadratic Kappa Score of the models are 0.931, 0.956 and 0.926 respectively.