An AI system can produce magnetic resonance imaging (MRI) scans with only a quarter of the data normally required, which could speed up the scanning process.
MRIs are created by placing a person inside a machine that scans the body and are often used to image brain development or muscle and tissue injuries. The speedy AI-based system, called FastMRI, was developed by researchers at Facebook AI and NYU Langone. It was trained on thousands of images gathered from 242 people.
The team then used the new system to create MRI scans of 108 people’s knees using 75 per cent less data taken during the scan to generate the finished image. The AI reconstruction uses less actual data, resulting in less time in an MRI machine. It manages this by filling in the “gaps” based on the images it was trained on.
Advertisement
The team then gave the 108 FastMRI scans to six radiologists, five of whom couldn’t distinguish them from MRI scans obtained using the traditional method.
“If the images look better than our standard images, and perform as well in a standard diagnosis, there’s no reason this can’t be part of the clinical routine very quickly,” says Michael Recht of NYU Langone in New York, one of the authors of the paper.
“Reducing the time it takes to do an MRI is a holy grail in radiology research,” says Luke Oakden-Rayner, director of medical imaging research at the Royal Adelaide Hospital, Australia. An MRI scan can require a patient to be in the machine for up to an hour; the FastMRI alternative ranged from between 8 and 11 minutes. “The results are very promising.” However, taking a quarter of the data doesn’t mean the time per scan is reduced to a quarter: preparing the patient still takes the same time.
Oakden-Rayner also cautions the research only looked at large injuries, whereas smaller issues could be overlooked by the AI, potentially falsely reassuring radiologists and leading to inaccurate diagnoses.
The authors say that all MRI scans produce images that are subjectively analysed. In their small experiment, they found that swapping out FastMRI images for traditional MRI ones would result in radiologists suggesting different clinical opinions no more than 4 per cent of the time. They are also testing the FastMRI model in other institutions to ensure similar results before seeking regulatory approval. “We’re not saying, based on this one study, everyone should switch what they do,” says Recht. “But we’re saying this is incredibly exciting.”
The initial findings are a basis to try other areas of the body. Work is under way on speedier MRI brain scans and abdomen scans – a trickier proposition because the soft tissue inside is always moving. The hope is that by reducing the time MRI scans take, the machines could be used more often, including in strokes, giving doctors the chance to learn more about patients with complicated issues.
American Journal of Roentgenology
