NV-Raw2Insights-MRI

Universal MRI reconstruction from undersampled k-space. A single model handles diverse protocols, anatomies, contrasts, and acceleration factors without task-specific fine-tuning.

Overview

NV-Raw2Insights-MRI is built on the Scalable Deep Unrolled Model (SDUM) framework. It combines a Restormer-based cascaded unrolled architecture with learned coil sensitivity estimation, sampling-aware weighted data consistency, and universal conditioning on protocol metadata. Trained on heterogeneous data from CMRxRecon2024, CMRxRecon2025, and fastMRI brain datasets, a single model achieves state-of-the-art results across cardiac, brain, and knee MRI reconstruction.

This project was conducted by NVIDIA in collaboration with the CMRxRecon Team, Fudan University, and Johns Hopkins University.

News

• [March 2026] — Released NV-Raw2Insights-MRI as part of the NVIDIA MedTech Open Models
• [February 2026] — Achieved 1st place across all four tracks in the CMRxRecon2025 Challenge without task-specific fine-tuning

Model Variants

Model	Cascades	Parameters	HuggingFace	License
NV-Raw2Insights-MRI-Small	6	230M	Download	NVIDIA Open Model
NV-Raw2Insights-MRI-Base	18	760M	Download	NVIDIA Open Model
NV-Raw2Insights-MRI-Large	34	1.4B	Download	NVIDIA Open Model

Checkpoints are automatically downloaded from HuggingFace when not provided locally.

Quick Start

Installation

pip install -r requirements.txt

Inference

python scripts/inference.py \
  -c configs/nv_raw2insights_mri_base.json \
  -i example \
  -o outputs/example_output_base

For multi-GPU inference:

torchrun --nproc_per_node=8 scripts/inference.py \
  -c configs/nv_raw2insights_mri_base.json \
  -i /path/to/input \
  -o /path/to/output

Documentation

Guide	Description
Setup	Full installation guide
Inference	Inference options, configs, multi-GPU
Training	Training and fine-tuning guide

Performance

Model Scaling (PSNR vs cascade depth)

Cascades (T)	PSNR (dB)	Parameters
1	28.73	42M
3	30.21	126M
6	32.09	253M
10	32.54	422M
18	33.18	759M

Inference Compute (per slice, NVIDIA H100, T=18)

Input Size	Time (s)	Memory (GB)
128x128	0.32	4.78
256x256	1.03	6.07
256x512	2.06	7.98
328x512	2.67	9.26
328x640	3.30	9.62
328x768	3.97	10.83

License

Component	License
Source code	Apache 2.0
Model weights	NVIDIA Open Model License

This project will download and install additional third-party open source software projects. Review the license terms of these open source projects before use.

Citation

@article{wang2025sdum,
  title={SDUM: A Scalable Deep Unrolled Model for Universal MRI Reconstruction},
  author={Wang, Puyang and Guo, Pengfei and Chai, Keyi and Zhou, Jinyuan and Xu, Daguang and Jiang, Shanshan},
  journal={arXiv preprint arXiv:2512.17137},
  year={2025}
}

Please also cite the CMRxRecon dataset papers.

Resources

• SDUM Paper — arXiv
• HuggingFace Model — Weights and model card
• CMRxRecon2025 Challenge — Benchmark