USC's MosaicMRI Doubles Open MRI Data With 2,671 Scans Across 4 Body Parts
A new open-source dataset of 2,671 raw musculoskeletal MRI volumes exposes how anatomy diversity, not just data volume, drives reconstruction models.
ProTakeaways
- MosaicMRI releases 2,671 raw multi-coil MSK MRI volumes and 80,156 slices, roughly 2x the largest prior public dataset.
- Covers spine, knee, hip, shoulder, pelvis, ankle and more with varied contrasts, orientations and coil counts.
- Stored in HDF5 with fastMRI-compatible layout, so existing pipelines run with minimal modification.
- E2E-VarNet baseline reaches 40.62 dB PSNR with 12 cascades trained on the full dataset.
- Training on mixed anatomies boosts PSNR by up to 6 dB in low-data regimes versus anatomy-specific models.
- Dataset and code available at mosaicmri.ai and github.com/AIF4S/mosaicmri.
Deep learning for MRI has been quietly stuck in a rut. Nearly every headline result on accelerated reconstruction, artifact removal, or segmentation traces back to the same handful of public benchmarks focused on brains and knees. A new dataset out of USC wants to break that habit by handing researchers the largest open collection of raw musculoskeletal (MSK) scans yet, along with a benchmark suggesting that what you train on matters more than how much of it you have.
MosaicMRI is a curated release of 2,671 fully sampled multi-coil volumes and 80,156 slices spanning spine, knee, hip, and ankle, with variation across orientations, contrasts (PD, T1, T2), and coil configurations. By both volume and slice count, it is roughly twice the size of the largest previously available public MSK MRI dataset.
Why brain and knee benchmarks held things back
To understand the gap, it helps to know what raw MRI data actually looks like. Scanners collect measurements in the frequency domain, called k-space, and images are reconstructed from that raw signal. Accelerated MRI reconstruction, the task of recovering a clean image from deliberately undersampled k-space, is the dominant benchmark for ML in this space, and progress has been shaped by whatever anatomies happened to be public.
That matters clinically. MSK MRI covers a diverse set of joints including the spine, shoulder, hip, knee, ankle, and wrist, each with distinct tissue composition and clinical indications. Imaging protocols vary substantially across regions in field of view, spatial resolution, coil selection, and motion susceptibility. A model tuned on knee cartilage acquisitions has no obvious reason to work on lumbar spine or shoulder tendons, and until now nobody could really test that at scale.