Blindness Risk Prediction caused by Diabetic Retinopathy from Retinal Image

Type-II diabetes is growing at an alarming rate worldwide. Eventually, it leads to visual impairment by damaging the retinal blood vessels, which is known as Diabetic Retinopathy (DR). A plethora of research is ongoing to automate the early detection of DR to help the doctor with periodic eye examinations due to advancements in AI based computer vision algorithms and camera technology. In this paper, we have proposed a comparison between two widely used very deep convolutional neural network architectures (ResNet-101 v2, InceptionResNet V2) with the comparatively new optimized and composite scalable architecture (EfficientNet B5) with while training them from scratch and using them as a pre-trained to transfer learning. We have found out the pre-trained EfficientNet B5 has outperformed our other candidates as well as available methods in the current literature by achieving accuracy of 97.78%. We also provide detailed enough information to make the result reproducible.