Abstract
Computationally synthesized blood vessels can be used for training andevaluation of medical image analysis applications. We propose a deep generativemodel to synthesize blood vessel geometries, with an application to coronaryarteries in cardiac CT angiography (CCTA). In the proposed method, a Wasserstein generative adversarial network (GAN)consisting of a generator and a discriminator network is trained. While thegenerator tries to synthesize realistic blood vessel geometries, thediscriminator tries to distinguish synthesized geometries from those of realblood vessels. Both real and synthesized blood vessel geometries areparametrized as 1D signals based on the central vessel axis. The generator canoptionally be provided with an attribute vector to synthesize vessels withparticular characteristics. The GAN was optimized using a reference database with parametrizations of4,412 real coronary artery geometries extracted from CCTA scans. Aftertraining, plausible coronary artery geometries could be synthesized based onrandom vectors sampled from a latent space. A qualitative analysis showedstrong similarities between real and synthesized coronary arteries. A detailedanalysis of the latent space showed that the diversity present in coronaryartery anatomy was accurately captured by the generator. Results show that Wasserstein generative adversarial networks can be used tosynthesize blood vessel geometries.