Even though MRIs are one of the best tools for diagnosing disease and injury, it takes a significant amount of time for the scanner to gather the necessary data. Staying completely still for an hour inside a loud metal tube can be a harrowing experience for patients in pain, and can be extremely difficult for children, leading to many failed scans. The time it takes to complete a single MRI scan also limits how many people can be scanned in a given day.
With fastMRI, artificial intelligence can generate accurate and detailed MRIs using only a quarter of the raw data that’s traditionally required for a full MRI. Because less data is needed, patients can spend far less time in the machine.
When I was at NYU, I saw that AI-based reconstruction methods had the potential to revolutionize medical image reconstruction. At that point, I was convinced that this was a technology to work on for the future.
MRIs are one of safest and most versatile imaging modalities we have, but their applicability is limited by their long scan times. In order to get a good scan, you need to lay still in the machine for several minutes which is difficult for patients who are too ill, or for little children. Reducing the scan time has the potential to make MRIs more accessible, improve patient comfort, and reduce medical costs.
Our researchers understand how grueling it can be for patients to stay completely immobile while a machine whirs and creaks around them. The shorter the time it takes for scans to complete, the better, especially for people who are extremely ill, very young, or suffer from anxiety and claustrophobia. Faster scanning can also increase the number of patients each machine can serve, drastically shortening wait times. These benefits are why our AI experts partnered with NYU’s MRI experts to not only develop fastMRI, but also publish the data, models, and code so that other researchers can build on the work and contribute to better patient care.
The fastMRI team built a neural network and trained it using NYU Langone Health’s open source dataset of knee MRIs, with NYU Langone ensuring that all scans were de-identified. By removing three-fourths of the raw data in each scan and then feeding the remaining information into the AI model, the model learned to generate complete images from the limited data. These images produced by the AI model didn’t just look like generic MRIs, but matched the ground truth image created by the standard slow MRI process. Radiologists and clinicians would be able to use these AI-generated images as they would traditional images. The only difference would be that the patient spent less time in the machine.
In a retrospective clinical study published in the American Journal of Roentgenology, radiologists reviewed both traditional MRIs and images generated with the AI model using 75 percent less raw data. The radiologists produced the same diagnoses with both. They also could not tell which images were created using the AI model. The study indicates that fastMRI images are interchangeable with those of regular MRIs, producing images that are just as accurate, useful, and reliable as those from a standard MRI.
Jure Zbontar, Florian Knoll, Anuroop Sriram, Tullie Murrell, Zhengnan Huang, Matthew J. Muckley, Aaron Defazio, Ruben Stern, Patricia Johnson, Mary Bruno, Marc Parente, Krzysztof J. Geras, Joe Katsnelson, Hersh Chandarana, Zizhao Zhang, Michal Drozdzal, Adriana Romero, Michael Rabbat, Pascal Vincent, Nafissa Yakubova, James Pinkerton, Duo Wang, Erich Owens, C. Lawrence Zitnick, Michael P. Recht, Daniel K. Sodickson, Yvonne W. Lui
Through our vision and research in ML, we developed wav2vec, a way to build speech recognition systems that require no transcribed data.
Facebook AI and Carnegie Mellon University’s Department of Chemical Engineering have joined to collaborate on the Open Catalyst Project.