Automated Grading of Diabetic Retinopathy with Ultra-Widefield Fluorescein Angiography and Deep Learning

Figure 1

Joint optimization of CycleGAN and CNN classifier. A fixed CycleGAN with a pair of extended classifiers was designed to analyze the DR-related pathological characteristics. We defined the domain as comprised of normal images and the domain as containing abnormal images. The generators and mapped and functions, respectively. When given a real domain image as the input, the cycle consistency loss constrains the output of to be consistent with the input, while a discriminator and a classifier cooperate to optimize the output of to approach domain . When the input is a real image of domain , the outputs of and are opposite to the input of domain .