MosaicMRI: A Diverse Dataset and Benchmark for Raw Musculoskeletal MRI

MosaicMRI is a large, diverse collection of fully sampled raw musculoskeletal MRI measurements for training and evaluating machine-learning methods. It spans multiple anatomies, orientations, imaging contrasts, and coil configurations. The work establishes VarNet reconstruction baselines and studies scaling with model capacity and dataset size, cross-anatomy transfer, robustness, and generalization under domain shifts. The results show that combining anatomies can substantially improve performance in low-sample regimes and reveal useful correlations across body regions.

Advances in Neural Information Processing Systems (NeurIPS 2026)