Transfer Learning Approach for Diabetic Retinopathy Detection using Efficient Network with 2 Phase Training

Diabetic Retinopathy is a diabetes complication that causes permanent blindness. This also affects both eyes, beginning with no visual symptoms. Yet it may lead to permanent blindness without adequate treatment. Early detection of diabetic retinopathy can be an opportunity to prevent vision loss. It involves many complicated and costly treatment methodologies and sophisticated analysis of retinal fundus images by expert doctors. Many researchers proposed different image processing and segmentation techniques. The unavailability of good quality of fundus images made the techniques unstable. Deep learning showed significant performance in various medical fields, including diabetic retinopathy. But due to a lack of high cost labeled sample dataset and performance issues due to the use of large parameters, made the approaches inefficient. In this paper, we have proposed an ensemble of 5 models using the optimal transfer learning model, EfficientNet-B5 with 2 phase training. We trained the model with a large number of sample datasets from Kaggle. Our model can early screen diabetic retinopathy and also achieved a high metric (quadratic weighted kappa score of 0.961).