Source-linked AI summary
fastMRI: An Open Dataset and Benchmarks for Accelerated MRI
Jure Zbontar, Florian Knoll, Anuroop Sriram, Tullie Murrell, Zhengnan Huang, Matthew J. Muckley, Aaron Defazio, Ruben Stern, Patricia Johnson, Mary Bruno, Marc Parente, Krzysztof J. Geras, Joe Katsnelson, Hersh Chandarana, Zizhao Zhang, Michal Drozdzal, Adriana Romero, Michael Rabbat, Pascal Vincent, Nafissa Yakubova, James Pinkerton, Duo Wang, Erich Owens, C. Lawrence Zitnick, Michael P. Recht, Daniel K. Sodickson, Yvonne W. Lui
TL;DR
MRI offers strong diagnostic imaging but often requires acquisition times exceeding 30 minutes, limiting throughput, comfort, and cost efficiency. This paper introduces fastMRI, a large-scale raw and clinical MRI dataset with standardized evaluation resources for machine-learning reconstruction, providing a foundation for advancing accelerated MRI research.
Problem
MRI acquisition can exceed 30 minutes, creating patient comfort, motion-artifact, throughput, and cost challenges that motivate accelerated reconstruction research.
Method
The paper releases raw k-space data, processed clinical MR images, evaluation metrics, baseline algorithms, and held-out test data through the fastMRI dataset and benchmark framework.
Results
The release provides 8,344 raw-MRI volumes, more than 1.57 million processed-image slices from 20,000 examinations, and standardized leaderboard evaluation using held-out test data.
Takeaways & Limitations
fastMRI connects data science and MRI research communities and offers raw k-space data intended to support reconstruction methods across scanners and imaging tasks.
Takeaways & Limitations
Extending challenge-developed methods to clinical use remains open because retrospective undersampling and baseline models do not explicitly account for spin relaxation, eddy currents, or field distortions.
Abstract
from arXiv · showhide
Accelerating Magnetic Resonance Imaging (MRI) by taking fewer measurements has the potential to reduce medical costs, minimize stress to patients and make MRI possible in applications where it is currently prohibitively slow or expensive. We introduce the fastMRI dataset, a large-scale collection of both raw MR measurements and clinical MR images, that can be used for training and evaluation of machine-learning approaches to MR image reconstruction. By introducing standardized evaluation criteria and a freely-accessible dataset, our goal is to help the community make rapid advances in the state of the art for MR image reconstruction. We also provide a self-contained introduction to MRI for machine learning researchers with no medical imaging background.
1 Introduction
MRI acquisition can exceed 30 minutes, creating cost, comfort, motion, and throughput problems, while reconstruction research lacks shared public data and evaluation standards. fastMRI addresses this gap with large-scale raw and processed MRI releases, benchmarks, and an MRI primer for machine-learning researchers.
- MRI scans can exceed 30 minutes, reducing throughput and patient comfort while increasing motion artifacts and exam costs.
- Parallel imaging and compressed sensing accelerate MRI by acquiring fewer or simultaneous measurements, but undersampling introduces artifacts requiring reconstruction methods.
- Machine-learning MRI reconstruction research lacks large-scale public datasets, consistent metrics, and baselines, hindering reproducibility and restricting participation.
- fastMRI releases raw data from 8,344 volumes and 167,375 slices, plus processed DICOM images from 20,000 knee and brain examinations totaling over 1.57 million slices.
- The paper provides an MRI acquisition and reconstruction primer alongside target tasks and benchmarks to connect data science and MRI research communities.
2 Introduction to MR Image Acquisition and Reconstruction
MRI reconstructs spatial images indirectly from frequency- and phase-encoded k-space measurements, with inverse Fourier transformation recovering an image estimate. Faster acquisition uses fewer samples or undersampling, while multi-coil measurements and learned or classical reconstruction methods address the resulting artifacts and noise.
- MRI measures frequency and phase responses in k-space rather than spatially resolved anatomy, using receiver coils and pulse sequences.
- The inverse multidimensional Fourier transform maps full k-space y to an image estimate ˆm, which remains noise-corrupted relative to the true image m.
- Undersampling can lower spatial resolution or create aliasing artifacts, so reconstruction strategies infer the underlying image structure from fewer k-space samples.
- Parallel MR Imaging: Each coil records Fourier components modulated by its position-dependent sensitivity map, providing multiple measurements of the imaged volume.
- Parallel MR Imaging: Parallel imaging can invert an undersampled multi-coil system when measurements suffice, but noise amplification and reduced effective rank limit practical acceleration.
- Classical reconstruction solves a regularized inverse problem, whereas machine-learning reconstruction learns a function from input-output pairs by minimizing expected or empirical loss.
3 Prior Public Datasets
Public medical-imaging datasets have advanced benchmarking, but prior MRI resources generally lack raw k-space data or remain too small for broad reconstruction research. fastMRI is positioned against this limited reference-data landscape.
- Public biomedical datasets and challenges have supported comparative evaluation and increased research activity in medical imaging.
- Prominent public MRI datasets generally do not include k-space imaging data.
- MR reconstruction lacks large-scale reference standards, with research relying on synthetic k-space data or small proprietary raw datasets.
- Existing public raw-k-space datasets remain small, and broader anatomy coverage is described as necessary to realize deep-learning potential.
4 The fastMRI Dataset and Associated Tasks
fastMRI provides raw multi-coil, emulated single-coil, ground-truth, and DICOM MRI data for standardized single-coil and multi-coil reconstruction tasks. Its splits, acquisition diversity, and reference reconstructions support training and evaluation, while DICOM data has important reconstruction and distribution differences.
- Dataset contents: The dataset contains raw multi-coil measurements, emulated single-coil data, ground-truth images, and DICOM images representing varied machines and settings.Ground-truth images are reconstructed from fully sampled multi-coil acquisitions, whereas DICOM images provide broader scanner and acquisition diversity.
- Associated tasks: The dataset supports single-coil and multi-coil reconstruction from undersampled k-space data, with each task approximating its associated ground-truth images.Official training and validation subsets contain fully sampled acquisitions; test and challenge subsets contain undersampled data without released ground truth.
- Dataset organization: Each volume is assigned to training, validation, test, or challenge subsets for model fitting, hyperparameter selection, and comparison across approaches.Test and challenge ground-truth reconstructions are withheld to reduce overfitting and enable evaluation through the benchmark process.
- Associated tasks: Single-coil reconstruction lowers the entry barrier for researchers and remains relevant to single-coil MRI machines still in use worldwide.The paper motivates simulated single-coil data both for accessibility and for practical relevance.
- Acquisition coverage: The raw data includes 1,594 fully sampled knee scans and 6,970 fully sampled brain MRIs acquired across multiple clinical scanners and field strengths.Knee scans use proton-density weighting with and without fat suppression; brain data includes axial T1, T2, and FLAIR images, with some T1 acquisitions using contrast agent.
- Dataset limitations: DICOM images broaden scanner and sequence diversity but do not directly correspond to measured raw data and may differ in quality from fully sampled ground truth.Vendor post-processing and parallel-imaging reconstruction affect DICOM images, and transfer learning may be necessary because their distribution differs from ground-truth images.
5 Metrics
The paper reviews standard pixel-wise, signal-to-noise, and structural metrics for MRI reconstruction, while noting that no single metric captures every aspect of image quality. It recommends familiar metrics as a starting point for developing improved evaluation measures.
- Common metrics: NMSE measures pixel-wise intensity differences between reconstructed and reference images, while PSNR measures image signal relative to distortion and noise.The paper recommends reporting NMSE over full image volumes rather than individual slices.
- Perceptual metrics: Perceptual metrics can compare images at pretrained-network representation levels or through downstream tasks such as classification.These approaches extend evaluation beyond pixel-wise error and local structural similarity.
- Metric recommendations: NMSE is recommended as the primary fastMRI reconstruction metric, but additional metrics are advised because NMSE can favor smoothness over sharpness.The paper expects fastMRI to support investigations into improved evaluation metrics and reconstruction algorithms.
- Structural metrics: SSIM evaluates structural similarity by measuring interdependencies among nearby pixels within sliding local windows.The implementation uses 7 × 7 windows with k1 = 0.01 and k2 = 0.03.
6 Baseline Models
The paper provides classical and deep-learning reconstruction baselines for single-coil and multi-coil accelerated MRI. The deep-learning baselines outperform classical methods, with larger models and more training data improving reconstruction quality, especially for harder acceleration settings.
- Baseline models: The dataset includes classical and deep-learning reference approaches tailored to both single-coil and multi-coil reconstruction.The classical methods derive from established MRI reconstruction techniques, while the deep-learning models are deliberately rudimentary starting points.
- Single-coil classical baseline: Single-coil reconstruction uses compressed-sensing formulations that enforce agreement with measured k-space while promoting sparsity through regularization.Common regularizers include L1, total variation, and wavelet penalties; the provided classical baseline uses total variation from BART.
- Single-coil classical baseline: Total-variation regularization can remove image detail, whereas wavelet regularization promotes sparsity in a discrete wavelet basis.The single-coil classical baseline runs 200 optimization iterations independently on each slice.
- Single-coil deep-learning baseline: 40-50% relative NMSE improvement is achieved by the best single-coil U-Net models over the classical methods on validation data.The U-Net models perform significantly better than the classical baseline.
- Model capacity: For 8-fold acceleration, the largest model performs significantly better than smaller models, whereas improvements saturate after 50 million parameters for 4-fold acceleration.Even the largest model, with over 200 million parameters, does not overfit the training data.
- Multi-coil deep-learning baseline: Multi-coil U-Net baselines substantially outperform classical baselines despite not using coil-sensitivity information or explicit parallel-imaging and sparsity modeling.Performance improves with model size, and the largest U-Net provides the best reported performance.
7 Discussion
MR reconstruction is an ill-posed inverse problem where visually plausible errors can mislead clinical interpretation, while current metrics and retrospective models leave important limitations unresolved.
- MR reconstruction recovers high-dimensional images from lower-dimensional measurements, making the inverse problem ill-posed because many images can match the same measurements.
- Clinical reconstruction must avoid plausible but incorrect images because radiologists use MRI to detect abnormalities, diagnose conditions, and recommend treatment.
- Global metrics such as NMSE, SSIM, and PSNR may not capture the detail needed to evaluate MRI reconstructions properly.
- Meaningful evaluation ultimately requires clinical studies testing whether fully sampled and undersampled reconstructions lead to the same diagnosis.
- Retrospective undersampling and baseline models do not explicitly account for spin relaxation, eddy currents, or field distortions.
- Extending challenge-developed methods to clinical use remains open because these physical effects depend on the object, scanner, and sampling pattern.
8 Conclusion
fastMRI provides a large public raw-MRI dataset with evaluation metrics, baseline algorithms, and held-out-test leaderboard evaluation. The authors position it as a benchmark intended to accelerate MR reconstruction research.
- fastMRI is described as the largest raw MRI dataset publicly available at the time, including raw k-space data for developing imaging-pipeline methods.
- The release includes evaluation metrics, baseline algorithms, and a leaderboard using held-out test data for consistent assessment.
- The dataset is intended to accelerate MR image reconstruction research and serve as a benchmark during algorithm training and validation.
10 Changelog
The changelog records the staged expansion of fastMRI from a knee-focused dataset to one that also includes neuro data.
- V1, released in November 2018, introduced the arXiv preprint describing the fastMRI knee dataset.
- V2, released in December 2019, added the neuro dataset.
A Raw k-space File Descriptions
Raw fastMRI k-space files use one HDF5 file per volume and store acquisition, scanner, patient, target-volume, and undersampling metadata alongside the k-space tensor.
- ISMRMRD files are converted into simpler HDF5 files that store each volume’s entire k-space in a single tensor.
- The acquisition attribute identifies knee protocols as CORPD or CORPDF and brain protocols as AXFLAIR, AXT1, AXT1POST, or AXT2.
- The ISMRMRD XML header preserves scanner, field-of-view, k-space-dimension, and sequence-parameter information.
- Each examination has an anonymized patient identifier, while training and validation files include target-volume norm and maximum-value attributes.
- The acceleration attribute records an undersampling factor of either 4 or 8.
- The file descriptions cover separate multi-coil and single-coil tracks, and some metadata attributes are restricted to test or training and validation datasets.
A.1 Multi-coil Track
The track packages training, validation, and test data for multi-coil and single-coil reconstruction, with k-space, sampling masks, and reference reconstructions organized in HDF5 files.
- Multi-coil Track: The multi-coil track provides separate training, validation, and test datasets with consistent HDF5 structures.Training and validation files contain documented tensor structures; test files are provided for evaluation.
- Reference Reconstructions: The multi-coil reference reconstruction is a root-sum-of-squares image, with knee images cropped to 320 x 320.Its tensor shape is organized by slices, height, and width.
- Multi-coil Track: Multi-coil k-space tensors use dimensions for slices, coils, height, and width.The undersampled training k-space and multi-coil k-space data share this organization.
- Sampling Masks: The sampling mask defines the undersampled Cartesian k-space trajectory, with one element for each k-space width position.This mask representation is specified for the multi-coil and single-coil tracks.
- Single-coil Track: Single-coil files contain emulated k-space plus root-sum-of-squares and inverse-Fourier-transform reconstructions of the cropped data.The reconstruction tensors have shape (number of slices, 320, 320).
B Classical Reconstruction with BART
BART supplies reproducible classical reconstruction baselines through coil-sensitivity estimation followed by regularized reconstruction. Its workflow also specifies file formats, k-space centering, and key command-line settings.
- Classical Reconstruction with BART: BART implements classical parallel-imaging and compressed-sensing reconstruction methods used to produce baseline MSE estimates and Figure 2 illustrations.The paper uses BART as a reproducible reference implementation for undersampled MRI reconstruction.
- Data Format: BART stores each multidimensional array as paired .hdr and .cfl files, with dimensions in ASCII and complex-float data in column-major order.Missing k-space values are represented by zeros, and Python and MATLAB interfaces support file access.
- K-space Representation: Centered k-space places low frequencies at the image center and high frequencies at the edges, while fftshift converts uncentered data.The documented bitmask 7 selects axes 0, 1, and 2 for the conversion command.
- Reconstruction Workflow: The parallel MRI workflow first estimates coil-sensitivity maps, then reconstructs the image assuming those maps are exact.BART provides this approach through the ecalib and pics commands, including ESPIRiT-based sensitivity estimation.
- Sensitivity Estimation: The central fully sampled reference region is used to estimate coil sensitivities and serves as the auto-calibration region.The masking procedure expresses the number of reference lines as a percentage of k-space width.
- Regularized Reconstruction: TV reconstruction uses regularization strength 0.05 and 200 optimization iterations in the documented command configuration.The command specifies TV regularization through the T option and an iteration count through -i200.