Dataset
Welcome to the official page for MR-RATE, a pioneering vision-language model and 3D medical imaging dataset that pairs textual reports with brain and spine MRI volumes. Following the approach of CT-RATE, the first 3D medical imaging dataset to pair images with textual reports, MR-RATE offers brain and spine MRI volumes matched with corresponding radiology reports and metadata along with derived registrations, segmentations, and pathology labels, all freely accessible to researchers.
This section will be updated with new releases, corrections, and derivatives.
A major challenge in computational research on 3D medical imaging is the lack of comprehensive datasets. Addressing this issue, we present the MR-RATE dataset, consisting of 705,254 non-contrast and contrast-enhanced brain and spine MRI volumes from 98,334 imaging studies of 83,425 unique patients, paired with corresponding radiology reports and metadata. Brain and spine MRI examinations are acquired from patients using MRI scanners and organised into multiple imaging sequence categories, including T1-weighted, T2-weighted, FLAIR, SWI, and MRA. Each study is paired with associated metadata and a radiology report, which is produced by the radiologist during clinical interpretation. Additionally, co-registrations, atlas-registrations, multi-label segmentations, and classified pathology labes are provided as derivatives, allowing researchers to incorporate these resources directly into their own workflows. Together, these components make MR-RATE a comprehensive dataset for multimodal brain and spine MRI research.
All MRI volumes are provided in their native acquisition space. DICOM files are converted to NIfTI format using dcm2niix, as NIfTI is more convenient for research workflows and downstream processing. This conversion also allows DICOM metadata to be selectively curated, removing identifying information while preserving relevant acquisition and study details, which are then saved in CSV format. Using an adapted and parallelized version of the BrainLesion Suite preprocessing module, a brain mask is predicted for each brain volume with HD-BET, and defacing is then applied with Quickshear to remove identifiable facial features for patient anonymization. The predicted brain masks and defacing masks are released alongside the defaced volumes, enabling researchers to leverage these directly in their own workflows.
All radiology reports are anonymized to remove patient healthcare identifiers and then structured into clinical information, technique, findings, and impression sections through an iterative LLM-based pipeline using Qwen3.5-35B-A3B-FP8 model via vLLM. Resulting reports are stored in CSV format.
In addition to the native-space data, MR-RATE offers four sets of derivatives:
To support reproducible research, we provide patient-level data splits. The splits are used internally for our own experiments and are shared openly so that the research community can benchmark and compare methods under consistent conditions.
| Split | # Patients | # Studies | # Series |
|---|---|---|---|
| Train | 75,000 | 88,985 | 638,345 |
| Validation | 3,425 | 3,781 | 27,003 |
| Test | 5,000 | 5,568 | 39,906 |
| Total | 83,425 | 98,334 | 705,254 |
MR-RATE dataset is released across four Hugging Face repositories. Study folders are zipped to comply with Hugging Face's per-repository file count limits. Since zipping disables include/exclude pattern filtering for downloads, data for each study is split across multiple zip folders per space and distributed across Hugging Face repositories, allowing users to download only their preferred data subset while staying within file count limits.
| Repo | Size | Content |
|---|---|---|
| Forithmus/MR-RATE (this repo) | 8.1 TB | Defaced native-space MRI volumes with brain masks and defacing masks, radiology reports, metadata, pathology labels, and data splits |
| Forithmus/MR-RATE-coreg | 17.6 TB | Co-registered MRI volumes in coreg-space, where moving modalities are registered to the T1-weighted center modality and center modalities are copied from native-space; registration transforms; and brain masks and defacing masks for the center modalities copied from native-space |
| Forithmus/MR-RATE-atlas | 12.3 TB | Atlas-registered MRI volumes in atlas-space, where all modalities are registered to the MNI152 atlas; registration transforms; and brain masks and defacing masks for the center modalities in atlas-space |
| Forithmus/MR-RATE-nvseg-ctmr | 415 GB | Voxel-wise anatomical multi-label segmentations in native-space, including brain segmentations for center modalities and body segmentations for all modalities |
To explore, download, and work with the dataset:
Coming soon
The MR-RATE dataset was used by NVIDIA to train NV-Generate-MR-Brain, a 3D latent diffusion model capable of generating high-quality, high-resolution synthetic brain MRI volumes for T1, Fluid Attenuated Inversion Recovery (FLAIR), T2, and Susceptibility Weighted Imaging (SWI). This model is available under the permissive NVIDIA Open Model License, enabling researchers and developers to adapt it for their own downstream medical AI applications.
When using this dataset, please consider citing the following related papers:
Coming soon
This study was approved by the Clinical Research Ethics Committee at Istanbul Medipol University (E-10840098-772.02-6841, 27/10/2023). All MRI volumes, metadata, and radiology reports were fully anonymized prior to analysis to protect patient privacy.
We are committed to fostering innovation and collaboration in the research community. To this end, all elements of the MR-RATE dataset are released under a Creative Commons Attribution–NonCommercial–ShareAlike (CC BY-NC-SA) license.
This licensing framework ensures that our contributions can be freely used for non-commercial research purposes, while also encouraging contributions and modifications, provided that the original work is properly cited and any derivative works are shared under similar terms.
For commercial inquiries related to MR-RATE, please contact: contact@forithmus.com.
This project is conducted by Forithmus and the University of Zurich, in collaboration with NVIDIA and Istanbul Medipol University.
We are grateful to NVIDIA for their support, which made this work possible. We also sincerely thank Istanbul Medipol University for their support and for providing the data used in this project. High-performance computing resources were provided by NVIDIA and the University of Zurich ScienceCluster. We would also like to thank the following individuals from NVIDIA for their contributions to the development of MR-RATE: Marc Edgar, Daguang Xu, Dong Yang, Yucheng Tang, Can Zhao, Andriy Myronenko, and Pengfei Guo.
This collaboration represents an important step toward the long-term mission to make high-quality medical intelligence accessible worldwide.
