I’ve got to say I’m intrigued by the latest from Facebook, a company which has recently been outed as making questionable choices about data privacy. Despite the kerfuffle, or perhaps because of it, Facebook is investing in some face-saving data projects.
Most recently, Facebook has announced that it will collaborate with the NYU School of Medicine to see if it’s possible to speed up MRI scans. The partners hope to make MRI scans 10 times faster using AI technology.
The NYU professors, who are part of the Center for Advanced Imaging Innovation and Research, will be working with the Facebook Artificial Intelligence Research group. Facebook won’t be bringing any of its data to the table, but NYU will share its imaging dataset, which consists of 10,000 clinical cases and roughly 3 million images of the knee, brain and liver. All of the imaging data will be anonymized.
In taking up this effort, the researchers are addressing a tough problem. As things stand, MRI scanners work by gathering raw numerical data and turning that data into cross-sectional images of internal body structures. As with any other computing platform, crunching those numbers takes time, and the larger the dataset to be gathered, the longer the scan takes.
Unfortunately, long scan times can have clinical consequences. While some patients can cope with being in the scanner for extended periods, children, those with claustrophobia and others for whom lying down is painful might have trouble finishing the scanning session.
But if MRI scanning times can be minimized, more patients might be candidates for such scans. Not only that, physicians may be able to use MRI scans in place of X-ray and CT scans, both of which generate potentially harmful ionizing radiation.
Researchers hope to speed up the scanning process by modifying it using AI. They believe it may be possible to capture less data, speeding up the process substantially, while preserving or even enhancing the rich content gathered by an MRI machine. To do this, they will train artificial neural networks to recognize the underlying structure of the images and fill in visual information left out of the faster scanning process.
The NYU research team admits that meeting its goal will be very difficult. These neural networks would have to generate absolutely accurate images, and it’s not clear how possible this is as of yet. However, if the researchers can reconstruct high-value images in a new way, their work could have an impact on medicine as a whole.
