NV-Generate-CTMR

NVIDIA MAISI 3D latent-diffusion model for generating synthetic CT and MR-brain volumes — no input scan required. The CT task (rflow-ct) generates an image, optionally with a paired 132-class anatomy mask whose organs/tumours you control. The MR-brain task (rflow-mr-brain) generates a brain volume in a chosen contrast, or a co-registered multi-contrast set (cross-contrast). For synthetic dataset creation, augmentation of rare pathologies, and privacy-preserving data sharing.

NVIDIA MAISI-v2. The CT model (NV-Generate-CT) and brain-MR model (NV-Generate-MR-Brain) are trained on large multi-institutional CT and brain-MRI cohorts; the MR-brain v0 weights use the MR-RATE dataset. Adult human anatomy.

Research generation of synthetic 3D CT / brain-MR volumes (with paired CT masks) for dataset creation, augmentation of rare pathologies, and privacy-preserving sharing. Not a diagnostic device; synthetic images must not be used as real patient data.

Output quality depends strongly on field of view (size × spacing) — out-of-distribution FOVs produce unrealistic volumes even when accepted. CT masks (Path A) contain only the organs you select. MR cross-contrast is same-seed multi-sequence synthesis, not translation of a real input. Large Z (≥256) is under development for the 20 GB deployment.

~1–4 min per 512×512×128 volume on a 20 GB GPU (rflow, 30 steps); image+mask adds a 1000-step mask-generation pass. Scales with output size and number of volumes.

CT

human

.nii.gz