Deep learning slashes real-world MRI scan times
Accelerated MRI with AI image reconstruction nearly halved orthopedic scan times while maintaining or even improving image quality in a prospective study conducted at New York University and published Jan. 17 in Radiology [1].
Patricia Johnson, PhD, and colleagues achieved the feat after training a deep learning image-reconstruction model on 298 clinical knee MR images acquired on a 3-Tesla scanner.
Next, they had 170 patients who were referred for knee MRI undergo a conventional accelerated 3T knee MRI protocol run with and without their experimental deep learning protocol.
Six radiologists subspecialized in MSK assessed all images for overall quality, presence of artifacts, sharpness and signal-to-noise ratio.
The reviewers’ scores placed the AI-reconstructed images as diagnostically equivalent to the conventional MRIs for detecting abnormalities.
In fact, in some cases, the AI images brought back better ratings.
More to the central point of the study, the deep learning protocol had a mean scan time of 5 minutes 33 seconds ± 16 seconds.
The conventional protocol compared unfavorably, clocking a mean scan time of 9 minutes 56 seconds ± 19 seconds.
As for the real-world diagnostics, the 170 patients had a total of 615 detected pathologic features across 19 categories of abnormality. The most common tear was a medial meniscal tear, detected in 64 participants, and the most common chondral abnormality was in the patellar cartilage, detected in 85 participants, Johnson and co-authors report.
“This work demonstrates that deep learning reconstruction of accelerated knee MRI at 3T does work reliably in a real clinical setting with real patients,” they remark in their discussion. “Important future steps include correlating reader assessments of pathologic findings with a surgical ground truth and establishing robustness across a variety of scanners and vendors.”
In invited commentary, Frank Roemer, MD, of the University of Erlangen–Nuremberg in Germany suggests the NYU deep learning model is likely to change daily MSK radiology workflows [2].
“We will perform the same number of examinations we are doing today in a much shorter period or will perform many more examinations in the same period,” Roemer writes. “This will increase the demand for image interpretation and the need for experts to evaluate these examinations.”
Roemer adds:
Whether this task will be performed solely by musculoskeletal radiologists or by machine learning algorithms supervised by human experts will have to be shown—but either is possible.”
In a news release, NYU Langone Health notes the Johnson et al study is part of the fastMRI initiative established in 2018 by that institution and Meta AI Research (formerly Facebook Meta Platforms).
The release quotes Michael Recht, MD, co-senior author of the present study.
“Our new study translates the results from the earlier laboratory-based study and applies it to actual patients,” says Recht, radiology chair at NYU Grossman School of Medicine. “FastMRI has the potential to dramatically change how we do MRI and increase accessibility of MRI to more patients.”
Dave P. has worked in journalism, marketing and public relations for more than 30 years, frequently concentrating on hospitals, healthcare technology and Catholic communications. He has also specialized in fundraising communications, ghostwriting for CEOs of local, national and global charities, nonprofits and foundations.