Blood Vessel Geometry Synthesis using Generative Adversarial Networks

The content on this page accompanies the MIDL Amsterdam submission Blood Vessel Geometry Synthesis using Generative
Adversarial Networks
by Jelmer M. Wolterink, Tim Leiner and Ivana Isgum. The full paper submission can be found on OpenReview.net.

In this work we propose a generative model to synthesize geometries of blood vessels, in this case coronary arteries. Such geometries could be used to augment training data sets for discriminative machine learning methods. The following video shows random examples of real coronary arteries geometries, extracted in a data set of 50 cardiac CT angiography scans.

 

 

We train a generative adversarial network (GAN) to synthesize plausible coronary artery geometries based on a random input vector sampled from a latent space pz. The video below shows examples of synthesized geometries based on these random input vectors.

 

Walk through latent space

We found that vessels with similar geometry are located close to each other in the latent space pz and that sampling from different parts of the latent space will result in different kinds of coronary arteries. In the video below, we walk around between three points in the latent space (shown on the right). Different points in the latent space lead to different geometries (shown on the left), in which the vessel length, shape and orientation change.

 

Conditioning on vessel length

We also train a condition GAN, in which we provide an additional attribute vector to the generator containing specifications of the geometries that we wish to synthesize. In the videos below, we again move around through the latent space, but we provide an attribute vector that requires samples to be of length 50, 100, 150, 200 or 250 mm. Note that this constrains the synthetic vessel length, but still leaves enough space for diversity in shape and orientation.

50 mm

 

100 mm

 

150 mm

 

200 mm

 

250 mm