The chapter deals with the construction of generative models using Variational Autoencoder (VAE) and Generative Adversarial Neural Networks to synthesize new medical data. VAE is a synthesis of two complete neural networks: an encoder E and a generator G, as well as the latent space connecting them and enabling them to carry out random transformation and interpolation. Generative Adversarial Nets (GAN) in their turn are built on the principle of interaction between a generative model (generator G) and a discriminating model (discriminator D). When creating generator G (both VAE and GAN), its architecture of a neural network based on convolutional layers, with the application of the new deep learning framework Tensorflow-addons is used. As E and D encoders, respectively, the models of transfer learning, problem domain-image feature vector are used in the work. The comparison between them is made in the chapter and the most optimal model for solving the proposed problem is selected. The chapter presents the results of the research obtained on the basis of VAE and GAN implementation.