Facebook AI Research (FAIR) and the New York University (NYU) School of Medicine’s Center for Advanced Imaging Innovation and Research (CAI2R) announced today they are sharing a standardized set of AI tools and baselines and MRI data as part of their joint research project fastMRI. The aim is to leverage AI-driven image reconstruction to achieve a tenfold reduction in MRI scan times. This is the first large-scale MRI data set of its kind, and is expected to serve as a benchmark for future research in the field.
In 2016, the NYU School of Medicine research team first showed that machine learning could cut scan durations by generating complete MR images from partial data inputs. In the research, MRI devices collected a series of individual 2D spatial measurements (k-space data). From these less detailed initial scans, neural networks trained on a large amount of k-space data used an AI image reconstruction technique to generate complete images capable of indicating a tumor, ruptured blood vessel, or other key diagnostic feature.
Researchers concluded the process requires “much less measurement data to produce the image detail necessary for accurate detection of abnormalities.”
FastMRI‘s major mission is to improve diagnostic imaging technology, and eventually increase patients’ access to the life-saving technology. The FAIR and NYU School of Medicine research team is also providing baseline models for ML-based image reconstruction from k-space data subsampled at 4x and 8x scan accelerations. They have already seen promising preliminary results for accelerating MR imaging by up to four times.
The release of the largest open source database of MR data could play a huge role in advancing the organization and acceleration of MR reconstruction and related work. The NYU School of Medicine hopes to provide a benchmark-ready dataset that will help address the challenge of consistency which has troubled the new research field of MR reconstruction.
According to the announcement, the initial release includes approximately 1.5 million MRI images drawn from 10,000 scans, as well as raw measurement data from nearly 1,600 scans. The team stresses that, as is the case with other fastMRI project data, the new dataset was gathered in close cooperation with NYU Langone’s Internal Review Board. NYU has fully anonymized the dataset, including metadata and image content, “manually inspecting each and every Digital Imaging and Communications in Medicine (DICOM) image for unexpected protected health information.”
More information on the FAIR and NYU School of Medicine collaboration and the fastMRI dataset can be found at FBCode.
Author: Jessie Geng | Editor: Michael Sarazen