CardioMM: A Universal Reconstruction Foundation Model for Cardiovascular Imaging

CardioMM is the first universal reconstruction foundation model for multimodal cardiovascular magnetic resonance imaging (CMR). Trained on MMCMR-427K—the world's largest CMR k-space database—it enables high-quality image reconstruction under 8-24x accelerated scanning, aiming to address the clinical pain points of excessively long CMR scan times (30-60 minutes) and unstable image quality.

Cardiovascular disease (CVD) is a major global health threat. CMR has become the gold standard for diagnosis due to its non-invasiveness, lack of radiation, and rich tissue contrast, but it faces two core issues: long scan times (30-60 minutes per session) and unstable image quality caused by differences in equipment, protocols, and patients across institutions. Traditional acceleration methods (parallel imaging, compressed sensing) require specific protocol tuning, while existing deep learning solutions are mostly designed for specific scenarios and lack generalization capabilities across devices/protocols.

To train a universal model, the team built the MMCMR-427K database with the following features:

• Data scale: 427,465 sets of multi-coil k-space data (≈3.5TB), from 6120 scans and 1504 subjects
• International coverage: 13 international centers (4 public repositories + 9 clinical centers)
• Device diversity: 15 scanners (0.55T-7T, from 4 major manufacturers)
• Rich modalities: 12 CMR modalities (cine imaging, T1/T2 mapping, etc.)
• Wide disease coverage: 17 types of cardiovascular diseases, covering populations from Asia, Europe, and America This data diversity lays the foundation for the universal model.

The core innovations of CardioMM lie in combining semantic context understanding with physical information consistency:

1. Dynamic adaptive mechanism: Adjusts reconstruction strategies based on input features to adapt to different scanners, protocols, and patient characteristics
2. Physical information constraints: Explicitly introduces k-space data consistency constraints to enhance robustness under high undersampling (8×-24×)
3. Multimodal unified framework: A single architecture handles different CMR modalities without separate training, reducing deployment costs.

• Internal validation: Cross-center validation shows that image quality metrics (PSNR, SSIM) and visual assessments outperform existing methods
• Zero-shot generalization: Generates high-quality images on unseen external center data, demonstrating cross-device/population universality
• Clinical indicator preservation: Under 8×-24× acceleration, 11 cardiac phenotype indicators and 3 myocardial biomarkers are preserved, meeting diagnostic needs
• Downstream complementation: Fills the upstream gap and provides a reliable foundation for downstream tasks such as segmentation and phenotype analysis.

The team has open-sourced the complete training and evaluation code, including:

• Data preprocessing/organization scripts
• Distributed training workflow
• Reconstruction inference scripts
• Evaluation and visualization tools
• Automated analysis workflow It also provides a traditional SENSE reconstruction implementation for easy comparison of different technologies.

• Community impact: The CMRxRecon Challenge (2023-2025) attracted over 11,000 participants from 125 countries; MMCMR-427K drove NVIDIA's universal MRI reconstruction model
• Clinical prospects: Shortens scan time to 5-10 minutes, reduces patient discomfort/artifacts, improves device utilization, makes CMR more feasible in emergency/pediatric settings, and narrows quality gaps between institutions It is expected to become a foundational component of the next-generation CMR workflow and an important milestone in the application of universal models in medical imaging.