Mobile assisted diabetic retinopathy detection using deep neural network

Diabetic retinopathy (DR) is a major microvascular complication resulting from diabetes and continues to have a serious impact on global health systems. Globally about 95 million people suffer from DR. This paper focuses on detection aspects of a mobile application developed to perform DR screening in real time. The application is powered by a tensorflow deep neural network architecture that is trained and tested on 16,798 fundus images. These images are preprocessed to remove noise and prepare them to be fed into neural network. Preprocessing steps involve averaging all the images using a 5×5 filter to improve the quality of images and then these images are resized to 256×256 pixels. After preprocessing the input dataset is fed into the neural network. The convolutional neural network model used in this project is MobileNets, which is used for mobile devices. The neural network has 28 convolutional layers and after each layer there is batchnorm and ReLU nonlinear function except at the final layer. The output from last layer is a class label either DR or no DR. The final accuracy of the model is 73.3%. This model is optimized to work on mobile devices and does not require Internet connection to run.