(The Eurographics Association, 2019) Haarburger, Christoph; Horst, Nicolas; Truhn, Daniel; Broeckmann, Mirjam; Schrading, Simone; Kuhl, Christiane; Merhof, Dorit; Kozlíková, Barbora and Linsen, Lars and Vázquez, Pere-Pau and Lawonn, Kai and Raidou, Renata Georgia
Generative adversarial networks have been shown to alleviate the problem of limited training data for supervised learning problems in medical image computing. However, most generative models for medical images focus on image-to-image translation rather than de novo image synthesis. In many clinical applications, image acquisition is multiparametric, i.e. includes contrast-enchanced or diffusion-weighted imaging. We present a generative adversarial network that synthesizes a sequence of temporally consistent contrast-enhanced breast MR image patches. Performance is evaluated quantitatively using the Fréchet Inception Distance, achieving a minimum FID of 21.03. Moreover, a qualitative human reader test shows that even a radiologist cannot differentiate between real and fake images easily.