FastMRI is a joint research project between Facebook AI and NYU School of Medicine, a department of NYU Langone Health, which was created to investigate the use of AI to make magnetic resonance imaging (MRI) scans up to 10 times faster. Last year, we released the largest publicly available data set of raw MRI measurements, as well as open-source tools and baseline results to empower the larger AI and medical imaging research communities to help tackle this problem. To advance the state of the art as quickly as possible, we are also announcing the first ever fastMRI image reconstruction challenge based on this research, and releasing the challenge data set on September 5, 2019.
The goal of this work is to advance AI and make MRI technology available to more people by reducing the time required to obtain diagnostic-quality images. MRIs contain a lot of redundant information, which allows images to be compressed significantly without a substantial loss in quality. The current practice is to collect a complete set of raw measurements for each image, which is time-consuming. One approach to reducing image acquisition time is to collect a subset of raw measurements (which contains salient information about the subject) and then train a neural network to extrapolate a full MRI image from this subset. Collecting fewer raw measurements directly translates to shorter scan times for patients.
We are opening this up to the broader community to learn what other approaches might be viable. In general, the quality of output produced by a neural network improves when it is trained on more data, and the fastMRI data set is the largest publicly available data set of raw MRI measurements. We have already released fully anonymized raw data, image data sets, and public leaderboards on the fastMRI website where the larger research community can easily compare their work with the baselines and evaluate their reconstructions
The competition officially begins on September 5, 2019. Participants will use the fastMRI challenge data set to produce reconstructions of the knee MRI cases using their methods. Reconstructions can be single-coil track with 4x acceleration or multicoil tracks with 4x or 8x accelerations. Submissions with the highest Structural Similarity measure (SSIM) scores will be evaluated by a panel of radiologists from several different institutions, who will determine the winner for each track.
One representative from each of the three winning teams will be invited to present his or her work at the Medical Imaging Meets NeurIPS Workshop at NeurIPS 2019. A separate challenge leaderboard (with scores based on a holdout challenge data set) will also be made public at the workshop. To enter, participants should review the submission guidelines and submit their reconstructions at fastmri.org no later than September 19. To be eligible to present at NeurIPS, participants should include a one-page abstract for the workshop with their challenge entry.
Our hope is that this challenge will ensure the reproducibility of the work and accelerate adoption of the resulting methods in clinical practice. By boosting the speed of MRI scanners, we can make these devices accessible to a greater number of patients and expand access to this important diagnostic tool.
Research Engineering Manager
Applied Research Scientist