Machine learning approach looks to reduce MRI scan times, costs
Researchers contend that MRI can be accelerated through a compressed sensing approach that uses an unsupervised end-to-end learning framework.
Magnetic Resonance Imaging is a non-invasive and versatile technology, but MRI scans are expensive. Now, machine learning offers a potential solution to the costly procedure.
“It’s quite expensive—it’s about $2,500 per scan in the U.S. in a hospital setting,” says Mert Sabuncu, an assistant professor in the School of Electrical and Computer Engineering at Cornell University, whose focus is on implementing innovative tools for analyzing biomedical images.
According to Sabuncu, an important driver of MRI cost is scan time. He points out that the reason MRIs are so expensive are that they take a long time to acquire the scan—anywhere from 15 minutes to 90 minutes per scan.
However, Sabuncu contends that MRI can be accelerated through a compressed sensing approach that uses a novel unsupervised end-to-end learning framework.
“The whole motivation of this work is to reduce the scan time to make this technology more accessible and more affordable,” Sabuncu told last month’s Machine Learning for Health Care conference, held in Ann Arbor, Mich.
The machine learning method trains a convolutional neural network on a set of full-resolution MRI scans, which are retrospectively under-sampled and forwarded to an anti-aliasing model that computes a reconstruction—which, in turn, is compared with the input.
In experiments with brain and knee MRI scans, Sabuncu contends that he and his colleagues demonstrate that the optimized under-sampling pattern can yield significantly more accurate reconstructions compared with standard under-sampling schemes.
The machine learning method—called Learning-based Optimization of the Under-sampling PattErn (LOUPE)—was implemented by modifying a U-Net, a widely used convolutional neural network architecture.
“Even with an aggressive eight-fold acceleration rate, LOUPE’s reconstructions contained much of the anatomical detail that was missed by alternative masks and reconstruction methods,” states a journal paper, which was submitted in late July and is under review but is available in pre-print. “Our experiments also show how LOUPE yielded optimal under-sampling patterns that were significantly different for brain vs. knee MRI scans.”
“This is relatively new work—it’s not like a fully baked, mature project,” observes Sabuncu, who notes that the results of the experiments were presented in June at the Information Processing in Medical Imaging conference.
He adds that the code is freely available online here at GitHub.
Going forward, Sabuncu says a pilot phase is beginning to prospectively validate the machine learning method in a “real MRI scanner and showing that we can get good reconstructions out of this.”
“It’s quite expensive—it’s about $2,500 per scan in the U.S. in a hospital setting,” says Mert Sabuncu, an assistant professor in the School of Electrical and Computer Engineering at Cornell University, whose focus is on implementing innovative tools for analyzing biomedical images.
According to Sabuncu, an important driver of MRI cost is scan time. He points out that the reason MRIs are so expensive are that they take a long time to acquire the scan—anywhere from 15 minutes to 90 minutes per scan.
However, Sabuncu contends that MRI can be accelerated through a compressed sensing approach that uses a novel unsupervised end-to-end learning framework.
“The whole motivation of this work is to reduce the scan time to make this technology more accessible and more affordable,” Sabuncu told last month’s Machine Learning for Health Care conference, held in Ann Arbor, Mich.
The machine learning method trains a convolutional neural network on a set of full-resolution MRI scans, which are retrospectively under-sampled and forwarded to an anti-aliasing model that computes a reconstruction—which, in turn, is compared with the input.
In experiments with brain and knee MRI scans, Sabuncu contends that he and his colleagues demonstrate that the optimized under-sampling pattern can yield significantly more accurate reconstructions compared with standard under-sampling schemes.
The machine learning method—called Learning-based Optimization of the Under-sampling PattErn (LOUPE)—was implemented by modifying a U-Net, a widely used convolutional neural network architecture.
“Even with an aggressive eight-fold acceleration rate, LOUPE’s reconstructions contained much of the anatomical detail that was missed by alternative masks and reconstruction methods,” states a journal paper, which was submitted in late July and is under review but is available in pre-print. “Our experiments also show how LOUPE yielded optimal under-sampling patterns that were significantly different for brain vs. knee MRI scans.”
“This is relatively new work—it’s not like a fully baked, mature project,” observes Sabuncu, who notes that the results of the experiments were presented in June at the Information Processing in Medical Imaging conference.
He adds that the code is freely available online here at GitHub.
Going forward, Sabuncu says a pilot phase is beginning to prospectively validate the machine learning method in a “real MRI scanner and showing that we can get good reconstructions out of this.”
More for you
Loading data for hdm_tax_topic #better-outcomes...