Fig. 2: CycleGAN principle and deep neural network architectures. | Nature Communications

Fig. 2: CycleGAN principle and deep neural network architectures.

From: Deep learning-enabled realistic virtual histology with ultraviolet photoacoustic remote sensing microscopy

Fig. 2

a Unsupervised training is performed with unpaired input data sets including virtual images X and true brightfield H&E-stained images Y. Forward and reverse generator transformations GH&E and GUV are trained concurrently with corresponding discriminators DH&E and DUV, which progressively improve their ability to classify generated synthetic images from true input examples. Images used for each cycle consistency and adversarial loss function are indicated. b Detailed architectures of the constituent deep convolutional neural networks used in the CycleGAN model. K convolution kernel size, S convolution stride.

Back to article page