Computer Vision-Aided Diabetic Retinopathy Detection Using Cloud-Deployed Deep Learning Framework

Diabetic Retinopathy is the major cause of blindness in diabetics, especially in underdeveloped countries. The use of deep learning algorithms has assisted in the early detection of Diabetic Retinopathy (DR). In this research, we propose a system that will help diagnose the DR quickly and efficiently while auto-training in the cloud based on the doctor’s feedback. This can be easily diagnosed using some easily available image acquisition systems such as smartphones’ camera and can be deployed in a low computational system in real-time. For the cloud training module, we analyzed five different transfer learning algorithms ensembled with a three-layered fully connected artificial neural network to create the feature map vectors followed by final classification. The pipeline gets auto-trained based on the kind of image that has been uploaded to the training module. We found that the system having DenseNet as the feature extraction performed a kappa score of 82.0% followed by the system having MobileNet and VGG-19 with 80.0 and 77.6% respectively. This trained module is then deployed for real-time analysis in AWS.