Abstract
BACKGROUND: Diabetic retinopathy (DR) produces bleeding, exudation, and new blood vessel formation conditions. DR can damage the retinal blood vessels and cause vision loss or even blindness. If DR is detected early, ophthalmologists can use lasers to create tiny burns around the retinal tears to inhibit bleeding and prevent the formation of new blood vessels, in order to prevent deterioration of the disease. The rapid improvement of deep learning has made image recognition an effective technology; it can avoid misjudgments caused by different doctors' evaluations and help doctors to predict the condition quickly. The aim of this paper is to adopt visualization and preprocessing in the ResNet-50 model to improve module calibration, to enable the model to predict DR accurately. RESULTS: This study compared the performance of the proposed method with other common CNNs models (Xception, AlexNet, VggNet-s, VggNet-16 and ResNet-50). In examining said models, the results alluded to an over-fitting phenomenon, and the outcome of the work demonstrates that the performance of the revised ResNet-50 (Train accuracy: 0.8395 and Test accuracy: 0.7432) is better than other common CNNs (that is, the revised structure of ResNet-50 could avoid the overfitting problem, decease the loss value, and reduce the fluctuation problem). CONCLUSIONS: This study proposed two approaches to designing the DR grading system: a standard operation procedure (SOP) for preprocessing the fundus image, and a revised structure of ResNet-50, including an adaptive learning rating to adjust the weight of layers, regularization and change the structure of ResNet-50, which was selected for its suitable features. It is worth noting that the purpose of this study was not to design the most accurate DR screening network, but to demonstrate the effect of the SOP of DR and the visualization of the revised ResNet-50 model. The results provided an insight to revise the structure of CNNs using the visualization tool.