What it does

fastMRI is a research project and dataset for accelerating MRI scans by acquiring fewer k-space measurements, then using machine learning to reconstruct the missing data. The repository provides PyTorch data loaders, subsampling masks, evaluation metrics, and reference implementations of baseline and published methods including U-Net and VarNet.

The interesting bit

The dataset includes fully anonymized raw k-space and DICOM data for knee, brain, and prostate scans — not just processed images. That means researchers can experiment with the actual measurement process, not just post-hoc image denoising. The project also tracks a long tail of published methods with reproducible code, from offset sampling tricks to adversarial banding removal.

Key highlights

• Datasets available for knee, brain, and prostate imaging via NYU Langone Health
• Reference models: U-Net, VarNet, adaptive VarNet, and classical baselines like ESPIRiT
• PyTorch Lightning modules for training and logging included
• Jupyter tutorial and demo training scripts to get started quickly
• MIT licensed; extensive publication list with linked implementations

Caveats

• The fastmri.org leaderboards are currently offline after domain transfer to NYU in April 2023; no timeline given for restoration
• Known memory leak with h5py from pip when converting to torch.Tensor; requires conda-installed h5py or HDF5 < 1.12.1

Verdict

Worth a look if you’re working on medical imaging reconstruction or need a well-curated, public MRI dataset with baseline code. Skip it if you need production-ready clinical deployment or a working leaderboard today.