fastMRI is a large-scale collaborative research project by Facebook AI Research (FAIR) and NYU Langone Health that explores how deep learning can accelerate magnetic resonance imaging (MRI) acquisition without compromising image quality. By enabling reconstruction of high-fidelity MR images from significantly fewer measurements, fastMRI aims to make MRI scanning faster, cheaper, and more accessible in clinical settings. The repository provides an open-source PyTorch framework with data loaders, subsampling utilities, reconstruction models, and evaluation metrics, supporting both research reproducibility and practical experimentation. It includes reference implementations for key MRI reconstruction architectures such as U-Net and Variational Networks (VarNet), along with example scripts for model training and evaluation using the PyTorch Lightning framework. The project also releases several fully anonymized public MRI datasets, including knee, brain, and prostate scans.
Features
- Open source PyTorch framework for accelerated MRI reconstruction research
- Includes MRI datasets (knee, brain, prostate) with raw k-space and DICOM data
- Provides U-Net, VarNet, and ESPIRiT-based reconstruction models
- Offers standardized evaluation metrics and reproducible training pipelines
- Compatible with PyTorch Lightning for modular training and logging
- Benchmarks from major publications and challenge datasets included
Project Samples
