A novel framework for severity level detection of diabetic retinopathy based on mobile edge computing and deep learning ensembles

In the current era of medical science, the detection of diabetic retinopathy in an automated way is very important, as it is one of the major reasons of vision loss in the middle-age population in the developed world. If it is detected at its early stages, it may be possible to alter the severe vision loss problem. Even though there are various approaches for feature extraction, the classification task for retinal images is still challenging for human experts. Utilization of internet-based intelligent systems may be useful for diagnosis of such health problems. Widespread usage of such systems requires maintaining a high detection accuracy as well as rapid response while utilizing the bandwidth efficiently. In this thesis, a framework is proposed for detecting severity levels diabetic retinopathy images using a smartphone as an edge computing device. The framework mainly consists of preprocessing, feature extraction, and classification steps, where the first step is performed on the edge device and the other two on the cloud computer. The overall performance is calculated based on the classification accuracy, response time, and total transmitted data. The preprocessing step involves cropping, resizing and unsharp masking. The classification is based on a concatenation ensemble of three benchmark convolutional neural network architectures that are EfficientNetB7, ResNet50, and VGG19. The proposed framework achieved a test accuracy of 0.96 and the edge computing improved the server response time and bandwidth occupation.