FastMRI initiative releases neuroimaging data set

December 12, 2019

Written byTullie Murrell, Anuroop Sriram, Nafissa Yakubova, Larry Zitnick

Written by

Tullie Murrell, Anuroop Sriram, Nafissa Yakubova, Larry Zitnick

Share

FastMRI, a joint research collaboration between Facebook AI and NYU Langone Health to use AI to speed up magnetic resonance imaging (MRI) scans, is announcing a new open source data set from NYU Langone Health, along with baseline models and a newly expanded research paper to help the AI research community accelerate and broaden research in this area.

Our research collaborators at NYU Langone Health are making available 6,970 fully de-identified cases of neuro MRIs in raw (k-space) format and 10,000 more cases containing 370,000 image slices in DICOM format as part of the fastMRI project. (Note that each case represents an imaging examination encompassing multiple sets of individual images.) Researchers can request access to the new brain MRI data set on NYU Langone Health’s fastMRI site, and the accompanying research paper and baseline models can be found here:

This is the largest public data set of raw k-space format brain MRIs available to researchers, and it follows our release last year of the largest knee MRI data set and the recently completed fastMRI image reconstruction challenge. K-space data is collected during scanning but typically discarded after it’s used to generate images. The information can be used to train models, validate their performance, and generally simulate how image reconstruction techniques would be used in real-world conditions. The fastMRI data set also includes undersampled versions of those measurements, with k-space lines retrospectively masked, to simulate accelerated partial-data scans.

“Brain imaging could benefit from rapid acquisition in so many ways. If you could take a 30-minute exam and accelerate it tenfold, that means a three-minute exam. For patients with neurological symptoms, you can imagine how that will have a huge impact in many ways: improving access to imaging, greater patient comfort, improved monitoring and safety for critically ill patients, better image quality in patients who have difficulty holding still. We hope to extend the work to fetal brain MRI as well, the one area that’s still difficult to evaluate using ultrasound,” said Yvonne Lui, M.D., Associate Chair of Artificial Intelligence and former Neuroradiology Section Chief in the NYU Langone Health Department of Radiology.

These examples from the fastMRI neuro data set show four different imaging sequence types: Axial T1, T1 post-contrast, fluid-attenuated inversion recovery (FLAIR), and T2.

Like all the data used or released by the fastMRI project, this data set was gathered as part of a study approved by NYU Langone Health’s Institutional Review Board, which oversees the responsible use of human research subjects and their data. NYU fully de-identified the data set, including metadata and image content, manually inspecting each and every image to ensure that no protected health information remained.

“The NIH recognizes the importance of sharing data sets to be able to rigorously develop and test new tools in order to advance science,” said Lui. “Here, we’ve made sure each image volume through the head has been cropped to exclude recognizable facial features, and every single image we’re releasing has been manually checked to make sure any protected health information has been removed.”

Using AI to aid radiologists

The fastMRI initiative aims to make scans up to 10 times faster than they are today, thereby improving the patient experience and making MRI scans less expensive and more accessible. Since MRIs are the most common imaging test used for the brain and spinal cord, the new neuro data set presents an important opportunity for researchers to test and refine their models. Brain MRIs are particularly challenging because movements of the eye and of fluid in the brain can introduce artifacts during image reconstruction. Faster scans that minimize the time in which these movements occur could help address this limitation.

The neuro data set will allow researchers to test their models with data from additional machine types, new sequence types, and different coil configurations that were not present in the previously released fastMRI knee data set. Radiologists also look for different diagnostic properties (such as contrast in texture between different neural tissue) in brain MRIs. These differences present an interesting and challenging machine learning problem to solve and will help researchers develop models that generalize to more clinical settings.

The fastMRI project has enabled researchers from both the medical imaging and AI communities to explore new approaches and compare their results. We hope this open, collaborative approach will accelerate AI research and improve access to potentially life-saving technology, and inspire more open and reproducible research practices in the field.

Written by

Research Engineer

Research Engineering Manager

Visiting Researcher

Research Scientist