Normalizing Flow for Synthetic Medical Images Generation

Abstract

Deep generative models, such as generative adversarial network (GAN) and variational autoencoder (VAE), have been utilized extensively for medical image generation. While these models made remarkable progress in medical image synthesis, they can not explicitly learn the probability density function of the input data and are highly sensitive to the hyperparameter selections. To mitigate these issues, a new type of deep generative model, called Normalizing Flows (NFs), have emerged in recent years. In this paper, we investigate NFs as an alternative for synthesizing medical images. In particular, we utilize realNVP, a popular NF model for the purpose of synthesizing medical images. To evaluate our synthesized images, we propose to utilize Wasserstien distance along with the permutation test to quantify the quality of the generated images. Within our quantifying metric, our results indicate that the two sample distributions, the first being the samples obtained from our NF model and second being the original dataset, are similar providing a promising indication of normalizing flow’s capability in medical images generation.