A deep-learning enhanced algorithm for the automated detection of diabetic retinopathy

A commonly occurring issue in diabetic patients, Diabetic Retinopathy (DR) does lead to progressive vision loss. It is not possible to restore the vision in the event of this abnormality's late-stage detection. Thus, this permanent condition's only solution will involve its early stage-identification as well as treatment. For sustaining a patient's vision, ophthalmologists will use their eyes' "fundus images," that is, images of their retina. Even so, the detection of a human eye abnormality via another naked human eye will tend to be quite costly as well as time-consuming, and in some cases, it can result in misjudgment because of the subjective difference as well as the considerations amongst the ophthalmologists. As a result, the DR detection is performed by means of the "deep learning" methodology which employs the fundus images. The introduction of this type of computer-based diagnosing system will aid in minimization of wrong diagnoses. Of late, deep learning techniques are commonly employed to accomplish accuracy amongst the image recognition or the feature detection systems for the classification as well as the regression. In this research, the Visual geometry group (VGG), the residual network (ResNet), and the Convolutional Neural Network (CNN) for DR recognition employing retinal images for training the neural network architecture to yield highly accurate results. The design of the CNN architecture is influenced by factors such as the nature of the input data, the complexity of the problem, and the desired performance. In this work, a customized CNN is presented that is specifically adapted to the DR detection problem. The custom 8-layer CNN's simulation results have demonstrated higher average sensitivity by 2.46, 1.42, and 4.99% for the VGG16, the ResNet-50, and the MobileNET, respectively.