Comparison of generative adversarial networks architectures for biomedical images synthesis

The article analyzes and compares the architectures of generativeadversarialnetworks. These networks are based on convolu-tional neural that widely used for classification problems. Convolutional require a lot training data to achieve desired accuracy. Generativeadversarialnetworks synthesis biomedical images in this work. Biomedi-cal medicine, especially oncology. For diagnosis oncology divided into three classes: cytological, histological, immunohistochemical. Initial samples very small. Getting trainingimages is challenging expensive process. A cytological datasetwas experiments. considers most common generative adversarialnetworks suchas Deep GAN (DCGAN), Wasserstein (WGAN),Wasserstein with gradient penalty (WGAN-GP), Boundary-seeking (BGAN), Boundary equilibrium (BEGAN). typical network architecture consists generator discriminator. discriminator CNN architecture.The algorithm deep learning image help ofgenerativeadversarialnet-worksis analyzed During experiments, following problems were solved. To increase initial number train-ingdata datasetapplied set affine transformations: mapping, paralleltransfer, shift, scaling, etc. Each was trainedfor certain numberof iterations. selected compared by timeand quality FID(FreshetInception Distance)metric. experiments implemented Python language.Pytorch as machine framework. Based softwarea prototype software module imageswas developed. Synthesis performed basis DCGAN, WGAN, WGAN-GP, BGAN, BEGAN architectures. Goog-le's online environment called Collaboratory experimentsusing Nvidia Tesla K80 graphics processor