[Paper Review] fastMRI: An Open Dataset and Benchmarks for Accelerated MRI

Jure Žbontar, Florian Knöll+25 more|arXiv (Cornell University)|Nov 21, 2018

Advanced MRI Techniques and Applications56 references572 citations

TL;DR

The paper introduces the fastMRI dataset, a large open collection of raw multi-coil k-space measurements and corresponding ground-truth images to benchmark machine-learning reconstruction methods for accelerated MRI, with standardized evaluation criteria.

ABSTRACT

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.

Motivation & Objective

• Provide a public, large-scale dataset of raw MR measurements and clinical images to accelerate research in ML-based MR reconstruction.
• Establish standardized evaluation criteria and baselines to enable fair comparison across reconstruction approaches.
• Introduce tasks (single-coil and multi-coil) and ground-truth references to guide model development and assessment.
• Offer an introductory resource on MRI concepts for ML researchers.
• Support the community in advancing state-of-the-art in accelerated MRI through open data and benchmarks.

Proposed method

• Release a large-scale dataset consisting of 8344 volumes (167,375 slices) of raw multi-coil k-space measurements and 20,000 knee and brain DICOM images.
• Define two reconstruction tasks: single-coil reconstruction approximating ground-truth from undersampled single-coil data, and multi-coil reconstruction from undersampled multi-coil data.
• Provide ground-truth references via root-sum-of-squares reconstructions and cropped central regions for evaluation.
• Introduce standardized evaluation metrics (NMSE, PSNR, SSIM, L1) and discuss their pros/cons for reconstruction quality assessment.
• Describe undersampling schemes (random for knee, equidistant for brain) and acceleration factors (4x or 8x) to simulate clinically relevant acceleration.

Experimental results

Research questions

• RQ1 How well can machine-learning models reconstruct MR images from undersampled k-space data across single-coil and multi-coil setups?
• RQ2 What evaluation metrics best capture reconstruction quality for accelerated MRI, and how do classical baselines compare to learning-based methods?
• RQ3 How does data diversity (knee vs brain, variable scanner strengths) affect reconstruction performance and generalization?
• RQ4 What role does ground-truth reference quality (RSS vs ESC) play in benchmarking ML reconstruction methods?

Key findings

• The dataset comprises 8344 volumes (167,375 slices) of raw multi-coil k-space data and 20,000 knee/brain DICOM images, enabling large-scale ML research in MR reconstruction.
• Two reconstruction tasks are supported: single-coil and multi-coil undersampled data, with corresponding ground-truth references and training/validation/test splits.
• Ground-truth images are generated via root-sum-of-squares reconstructions and cropped central regions, while DICOM data reflect broader scanner variability and post-processing differences.
• Undersampling is performed retrospectively with 4x or 8x acceleration, including central fully-sampled regions and variable remaining k-space coverage to facilitate CS-style benchmarks.
• The paper discusses NMSE, PSNR, SSIM, and L1 as evaluation metrics and notes the limitations and complementary insights each provides for reconstruction quality.
• Baseline methods include classical TV-based reconstructions and rudimentary deep-learning models such as a UNET, providing reference points for future improvements.