NV-Generate-MR-Brain

A brain MRI examination is never one image. T1-weighted volumes carry anatomy, FLAIR isolates white-matter lesions, T2 shows fluid and edema, and susceptibility-weighted imaging reveals microbleeds — and which of those a site acquires, in which plane, is a matter of local protocol. Anyone assembling a training cohort for brain segmentation or lesion detection therefore inherits a ragged matrix: thousands of one contrast, a handful of another, and no way to fill the holes short of scanning more patients.

NV-Generate-MR-Brain generates the missing volumes. It is the brain-specific checkpoint (rflow-mr-brain) of NVIDIA's NV-Generate-CTMR family, released in March 2026 by NVIDIA with the University of Zurich, Istanbul Medipol University and Forithmus. It carries over the rectified-flow latent generative recipe that MAISI-v2 introduced for CT — itself the successor to the 1000-step DDPM used by MAISI — and trains it on brain MRI alone.

That specialization is what separates it from its sibling NV-Generate-MR, which spans prostate, breast, abdomen and brain and whose documentation recommends fine-tuning on your own data for a new anatomy. Because every training volume here is a head, the brain checkpoint doubles the through-plane extent to 512x512x256 voxels at 0.45x0.45x0.7 mm, adds susceptibility-weighted imaging to the contrast vocabulary, extends skull-stripped output — which the general model offers for FLAIR alone — to every contrast, and is the variant NVIDIA's own guide tells brain users to prefer.

#Key Features

• Contrast and skull state as one integer: modality codes 9, 10, 11, 16 and 20 emit whole-brain T1w, T2w, FLAIR, MRA and SWI; codes 29 to 33 emit the skull-stripped counterparts. Switching output type is a config edit, not a second model.
• Thirty steps to a volume: rectified flow learns a near-straight velocity field, so sampling is an ODE solve of 30 steps rather than a long denoising chain, with classifier-free guidance applied to the modality label at scale 10.
• Field of view set per run: voxel dimensions times spacing define the anatomy that actually fits in frame, and both are inference arguments. The inference guide publishes recommended pairs per contrast and acquisition plane, drawn from the training distribution.
• Single-GPU generation: inference runs from the downloaded checkpoint on a 16 GB GPU under MONAI Core 1.5, writing NIfTI volumes; a hosted Space exercises the model with no GPU at all.

#Technical Details

The generator is a 240M-parameter 3D UNet with attention blocks operating in the latent space of a variational autoencoder shared with the CT checkpoints, so the diffusion model never touches full-resolution voxels. Unlike the paired ddpm-ct variant it takes no body-region input — the anatomy is fixed — and conditions only on the modality code. Training used the MR-RATE brain corpus assembled by Forithmus, the University of Zurich and Istanbul Medipol University: the model card reports roughly 28,000 scans for training with about 8,000 held out for testing and 4,000 for evaluation, while the repository's field-of-view tables tally 318,825 unique source volumes across contrast and plane, each preprocessed in both whole-brain and skull-stripped form. That 318,825 is a floor rather than a training-set size — the table covers axial, sagittal and coronal acquisitions only, and excludes the oblique scans the guide says the model also trained on. A third figure sits alongside both: NVIDIA's data documentation describes a successor checkpoint trained on 73,516 subjects and 665,371 scans, though the shipped configs and download script still fetch the v0 weights described here. Coverage is uneven by design of the source clinic — sagittal T1 and FLAIR dominate, while the guide flags sagittal and coronal SWI, and MRA in any plane, as too sparse for reliable output. NVIDIA has published no image-quality benchmark for this checkpoint; the repository's FID comparison covers the CT variants only.

#Applications

The immediate use is data augmentation where real brain MRI is scarce or unshareable: enlarging cohorts for segmentation and lesion-detection training, generating the contrast a retrospective study lacks, or producing volumes that carry no patient identity for distribution across institutions. The skull-stripped output feeds downstream tools that expect brain-extracted input without an extra preprocessing stage, and the per-plane field-of-view control makes it practical to stress-test an imaging pipeline against acquisition geometries a single site never produces.

#Impact

NV-Generate-MR-Brain extends an openly licensed synthetic-imaging family from CT into neuroimaging, and its weights carry the NVIDIA Open Model License, which permits commercial use — a materially different footing from the non-commercial terms on its general MR sibling. That permission covers the weights, not their provenance: MR-RATE itself is CC BY-NC-SA 4.0, and its publishers direct commercial inquiries to the dataset owners rather than granting them in the licence. The honest limits are worth stating: it is a version 0.1 research artifact, explicitly not a clinically validated device; the cross-modality synthesis the model card names as its primary use case (T1 to FLAIR and back) is not shipped, and unconditional single-contrast generation is what works today; and asking for a field of view outside the training distribution yields unusable volumes rather than a warning.