fastMRI: An Open Dataset and Benchmarks for Accelerated MRI (1811.08839v2)

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.

• The paper introduces the fastMRI dataset as a large-scale, open resource for benchmarking accelerated MRI reconstruction methods.
• It details both compressed sensing and deep learning techniques, with evaluations based on NMSE, PSNR, and SSIM metrics.
• The dataset aims to enhance clinical MRI efficiency by reducing scan times and standardizing evaluation protocols for improved patient care.

An Overview of the fastMRI Dataset and Benchmarks for Accelerated MRI

The paper "fastMRI: An Open Dataset and Benchmarks for Accelerated MRI" presents a significant contribution to the domain of medical imaging by introducing the fastMRI dataset. This dataset comprises raw MRI measurement data and clinical MRI images, curated to aid the development and benchmarking of machine learning algorithms for MRI reconstruction. The principal authors, Jure Zbontar, Florian Knoll, Anuroop Sriram, et al., have collaborated to provide a comprehensive resource pivotal for enhancing the state of the art in MRI acceleration.

Introduction and Motivation

Magnetic Resonance Imaging (MRI) is renowned for its superior soft tissue contrast, essential for diagnosing various conditions, including neurological, musculoskeletal, and oncological diseases. However, MRI suffers from prolonged acquisition times, often exceeding 30 minutes, which compromises patient comfort and throughput, thus escalating healthcare costs. The primary motivation behind this dataset is to facilitate the acceleration of MRI, thereby addressing these clinical inefficiencies.

Compressed Sensing and Machine Learning in MRI

In recent years, the introduction of Compressed Sensing (CS) and machine learning approaches has shown promise in reducing MRI scan times. CS techniques enable the acquisition of fewer measurements, thereby challenging the Nyquist-Shannon sampling theorem, but necessitating robust reconstruction algorithms to mitigate aliasing artifacts. Concurrently, machine learning approaches have emerged, leveraging large-scale datasets to innovate MRI reconstruction significantly. Despite these advancements, the field has lacked standardized datasets and benchmarks, hindering reproducibility and broad accessibility. The fastMRI dataset aims to fill this gap by providing a large-scale, publicly accessible dataset with clear evaluation metrics.

Dataset Composition

The fastMRI dataset includes:

1. Raw Multi-Coil K-Space Data: Unprocessed complex-valued multi-coil MRI measurements, critical for developing multi-coil MRI reconstruction algorithms.
2. Emulated Single-Coil K-Space Data: Simulated from multi-coil data to approximate single-coil acquisitions, enabling single-coil reconstruction algorithm evaluation.
3. Ground-Truth Images: Images reconstructed from fully-sampled k-space data.
4. DICOM Images: Processed images from a larger variety of MRI systems and settings, enhancing the dataset's diversity.

This data facilitates two primary reconstruction tasks: single-coil and multi-coil image reconstruction. It encompasses 1,594 knee scans and 6,970 brain scans, providing a substantial volume of data for training and validation.

Metrics for Evaluation

The paper outlines various metrics for evaluating MRI reconstruction quality, emphasizing:

• Normalized Mean Squared Error (NMSE)
• Peak Signal-to-Noise Ratio (PSNR)
• Structural Similarity Index (SSIM)

These metrics enable consistent and objective assessment of reconstruction algorithms.

Baseline Models

Two reference reconstruction approaches—classical and deep-learning methods—are provided as baselines:

• Classical Reconstruction: Based on compressed sensing, utilizing Total Variation (TV) regularization, implemented through the BART toolkit.
• Deep Learning Approaches: Centered on the U-Net architecture, trained end-to-end to minimize mean absolute error with respect to ground truth images.

Empirical results highlight the superior performance of U-Net models over classical methods, particularly for higher acceleration factors, demonstrating the potential of deep learning in MRI acceleration.

Implications and Future Directions

The fastMRI dataset's implications are multifaceted:

• Clinical Impact: Accelerated MRI can significantly enhance patient throughput and comfort, reduce motion artifacts, and decrease healthcare costs.
• Research Advancement: The dataset serves as a benchmark, fostering reproducible research and innovation in MRI reconstruction algorithms.
• Methodological Development: It encourages the exploration of new evaluation metrics and machine learning models, essential for further breakthroughs in the field.

The paper's introduction of fastMRI thus represents a critical step toward democratizing MRI research, promoting collaboration across the data science and medical imaging communities. Future developments might include integrating physical effects like spin relaxation and field distortions into the reconstruction models, extending the dataset's applicability in clinical settings.

In conclusion, the fastMRI dataset is a robust resource designed to catalyze advancements in MRI technology. By providing extensive raw data and standardized benchmarks, it empowers the research community to make significant strides in reducing MRI acquisition times, ultimately benefiting clinical practice and patient care.