Deep Learning Method Helps Correct Motion-Corrupted Brain MRIs

Researchers from the Massachusetts Institute of Technology (MIT) have developed a deep learning (DL) model that can correct magnetic resonance imaging (MRI) scans corrupted by motion artifacts, according to a recent arXiv preprint paper.

The researchers indicated that as an imaging modality, MRIs are superior to computed tomography (CT) scans or X-rays for capturing high-quality soft tissue contrast, but sacrifice speed. MRIs are also particularly sensitive to motion, even small movements like deep breaths.

These movements can obscure crucial details from the clinician reading the scan by creating image artifacts. The presence of these artifacts can corrupt a whole image, the authors noted, potentially putting patients at higher risk of misdiagnosis or inappropriate treatment course.

“Motion is a common problem in MRI,” said Nalini Singh, an Abdul Latif Jameel Clinic for Machine Learning in Health (Jameel Clinic)-affiliated PhD student in the Harvard-MIT Program in Health Sciences and Technology (HST) who served as lead author of the paper, in the press release. “It’s a pretty slow imaging modality.”

The type of imaging required will determine the length of an MRI session, but most take between a few minutes and an hour. As the scan’s timeframe lengthens, so do opportunities for small movements that could create artifacts in the resulting scan.

To minimize movement, clinicians may anesthetize a patient or ask that they limit their deep breathing. However, these options are not always feasible in patient populations susceptible to motion, including children and individuals with psychiatric disorders.

To combat this, the researchers worked to build a model capable of correcting corrupted MRIs. Their method is designed to computationally construct motion-free images using motion-corrupted data without altering the scanning procedure in any way.

“Our aim was to combine physics-based modeling and deep learning to get the best of both worlds,” Singh explained.

The method’s success hinges on the image output being consistent with the actual measurements of what is being depicted. Without this consistency, the model “hallucinates” to generate realistic-looking images that contain physical and spatial inaccuracies upon closer inspection.

These inaccuracies could negatively impact diagnoses and patient outcomes, the research team highlighted, meaning that ensuring that the model does not hallucinate is key.

The authors also noted that their model and similar technologies would not only improve patient outcomes, but also help tackle excess hospital expenditures related to repeated MRI imaging resulting from motion artifacts.

In the future, the researchers may examine how this approach could be applied to MRIs of other body parts or with more sophisticated head motions, such as fetal MRIs.

“This line of work from Singh and company is the next step in MRI motion correction,” said Daniel Moyer, PhD, an assistant professor of computer science at Vanderbilt University who did not participate in the research, but whose work focuses on machine learning (ML) in medical imaging. “Not only is it excellent research work, but I believe these methods will be used in all kinds of clinical cases: children and older folks who can't sit still in the scanner, pathologies which induce motion, studies of moving tissue, even healthy patients will move in the magnet... In the future, I think that it likely will be standard practice to process images with something directly descended from this research.”

Others are also looking to improve MRI through the use of advanced technologies.

In January, researchers from New York University (NYU) Grossman School of Medicine and Meta AI Research revealed that artificial intelligence (AI) could reconstruct MRIs faster and more accurately than traditional methods.

The work is part of the two organizations’ fastMRI initiative, which aims to use AI to improve the speed of MRI scans.

The AI model’s performance was approximately four times faster than that of standard methods, suggesting that such an approach could help reduce long patient wait times and expand MRI access.