In this work we set a baseline for using Generative Adversarial Networks (GAN) togenerate 3D medical images. Specifically, we trained a GAN to generate complete3D chest computed tomography (CT) scans up to 512x512x128 volumes. Most notably this is motivated by the purpose of anonymizing data for easier sharing and applications for low-dosage radiotherapy planning. Because of the memory and compute requirements, the model was trained on a CPU-based supercomputer. End-to-end model training took +-8 days. Additionally, we propose two metrics for large-resolution 3D volume comparison. In this paper, we report first results using these metrics, hoping to be bettered inthe future.To our knowledge, this is the largest-scale experiment of generative models so far,and the first to generate full-resolution complete 3D CT scans.
Generating CT-scans with 3D Generative Adversarial Networks Using a Supercomputer
D. Ruhe, V. Codreanu, C. van Leeuwen, D. Podareanu, V. Saletore and J. Teuwen
Medical Imaging meets NeurIPS 2019.