AICDR: Automatic and Improved Classification of Diabetic Retinopathy using deep Learning

Abstract

Diabetic retinopathy detection is time-consuming and challenging because resources and knowledge are needed to establish whether the disease is present. It is one of the main causes of blindness and is brought on by alterations in the retina's blood vessels. Diabetic retinopathy primarily results from diabetes. A manual diagnosis of diabetic retinopathy includes expert clinicians, resources, and specialized equipment. This process is costly and is not accessible wherever it is most needed. The goal of this research is to develop an automated, reliable, and cost-effective method for the high-accuracy identification of retinopathy. Physicians and patients can examine diabetic retinopathy more easily thanks to automated detection of the condition utilizing digital color retinal imaging. The retina extraction function can help determine the severity of the disease. This work uses multiple deep learning algorithms to automatically diagnose and categorize retinal pictures into five classes: severe non-proliferate DR, proliferation DR, mild DR, moderate DR, and no-DR. APTOS 2019 is a dataset used for deep convolutional neural network training. Accuracy was enhanced via transfer learning on deep convolutional neural networks that had previously been trained using ImageNet. In our study, the best deep learning models Resnet-18 and Resnet-50 achieved 88.2% and 92% accuracy on the validation range, generating 96% and 97% sensitivity on 100 test images from the same dataset, respectively. Results are described with the error rate, precision, sensitivity, and specificity using the confusion matrix, which describes errors explicitly in model misclassification.