For most of the past fifteen years, if you asked Li Feng, PhD, what he’s working on, the broad answer would have likely been the same: developing imaging techniques centered on radial MRI. But today, Dr. Feng, associate professor and director of rapid imaging in the Department of Radiology at NYU Grossman School of Medicine and imaging scientist with the Bernard and Irene Schwartz Center for Biomedical Imaging at NYU Langone Health, has his sight on a different horizon.
Opening to new directions, Dr. Feng’s lab is currently advancing quantitative liver MRI, denoising methods for low-field MRI, devising more anatomically accurate diffusion-weighted imaging techniques, and developing new ways for MRI scanners to become faster and more patient-specific by drawing on “memory” of prior exams.
The shift has been deliberate, and part of the thinking behind it appeared in a recent “mini review” authored by Dr. Feng in the journal Magnetic Resonance in Medicine. There, he frames radial approaches in a context where the value of ultimate clinical impact supervenes that of technical advancement.
“Why has radial MRI, despite decades of innovation and strong research enthusiasm, not evolved into a dominant clinical acquisition strategy?” asks Dr. Feng in the piece. Technical advancement has value,” he explained. Even so, in order to significantly shape day-to-day radiology, new technologies must not only offer innovation but also bring practical and generalizable benefits. “It wasn’t easy for me to write this,” said Dr. Feng in a recent interview. “We don’t have many reviews like that, trying to frame a topic from a different angle.”
The ‘paradox’ of radial sampling
In magnetic resonance imaging, raw data acquired by the scanner are recorded in a domain called k-space, which lives in a frequency dimension that corresponds to the physical space being imaged. In an overwhelming majority of clinical scans, data are written linearly: filling a row, then continuing in the next row, and so on, a bit like lines of text on a page. This acquisition method is called Cartesian sampling. Known for simplicity of reconstruction and high tolerance to scanner imperfections, Cartesian methods have been the backbone of MRI acquisition in radiology for decades.
But there are many more ways to scan. Data can be recorded in virtually any two- or three-dimensional arrangements, and k-space can be filled in varying order. Acquisition strategies can also be combined to integrate features of both Cartesian and non-Cartesian sampling.
Radial acquisitions are named after the shape of “spokes” or “blades” that record raw data in a pattern that resembles an asterisk. Imaging researchers often give these sampling trajectories geometrically evocative names, such as 3D koosh ball or PROPELLER.
In a radial acquisition, spokes cross each other in the central region of k-space—with two important consequences: One, spokes have data in common at their intersections; Two, each spoke samples an equally representative area of the imaged region. These characteristics, in turn, give rise to several desirable features. One is a self-centering quality, which imbues the method with natural robustness to patient motion. Another is a self-healing property in which the removal of individual spokes does not significantly degrade the overall acquisition. A third benefit is flexibility: radial data can be retrospectively sorted and reconstructed anew to obtain clearer images of physiological events of interest, such as the moment when a contrast agent perfuses an organ or a specific stage of the cardiac cycle.
The technical benefits have practical significance, especially for patients who are unable to perform breath holds or remain still during an MRI exam. For them, radial acquisitions can deliver better and more reliable images than can the conventional Cartesian methods.
But the very mechanism that underlies these upsides also entails several downsides. Reconstructing images from radial data, rather than from a row-by-row array, requires significantly more complex math. The process also tends to create distinctive “streaking” artifacts in accelerated scans and residual motion blurring, both of which lend to the images a somewhat soft visual quality. Technical MRI development always seeks a balance of multiple advantages and disadvantages, so the overarching research question has been whether advances in radial acquisitions can create techniques in which the benefits outweigh the drawbacks for clinical imaging. The answer—Dr. Feng contends in the mini review—has been both yes and no.
Mountains beyond mountains
Dr. Feng began working on radial MRI in 2011 as a second-year graduate student in biomedical imaging and technology at NYU Grossman School of Medicine. He and his colleagues and advisors set out to integrate radial sampling with a then recently introduced accelerative approach called compressed sensing and another, well established acceleration method called parallel imaging.
The project turned into a doctoral dissertation proposing GRASP MRI (short for golden-angle radial sparse parallel MRI). The technique would combine tolerance of patient-motion with the flexibility to retrospectively reconstruct selected data, and the scans would be fast enough for radiology clinics. These advantages contributed to GRASP becoming a kernel of a whole new vision for MRI, a paradigm in which the conventional stop-and-go concatenation of scans could one day be replaced by a single acquisition.
“It was a major component of CAI2R in the first NIH funding period,” said Dr. Feng, referring to the role GRASP played at the Center for Advanced Imaging Innovation and Research between 2014 and 2019. Just three years into the center’s first funding period, the method became a product on scanners manufactured by Siemens Healthineers, reaching the global medical MRI market under the name Compressed Sensing GRASP-VIBE. The translation validated the research center’s model of multidisciplinary and multi-sector collaboration aimed at bringing technical innovations to clinical practice in years rather than decades.
In the wake of this success, Dr. Feng continued to lead investigations extending the technique. Among these extensions: koosh-ball GRASP (adapted for four- and five-dimensional cardiac imaging), RACER-GRASP (guided by aortic contrast enhancement), GROG-GRASP (with improved computational efficiency), UTE-GRASP (adapted for lung imaging), GRASP-Pro and XD-GRASP Pro (with performance improvements), MP-GRASP and MP-Dixon-GRASP (adapted for fast T1 and fat/water separation). These advances spanned years during which Dr. Feng held consecutive primary affiliations with NYU Langone Health, Memorial Sloan Kettering Cancer Center, and Icahn School of Medicine at Mount Sinai, before returning in 2023 to NYU Grossman School of Medicine to lead research on rapid imaging.
Rather than feeling satisfaction with what amounts to an impressive portfolio by traditional measures of success in medical academic research, Dr. Feng invited reflection.
“Our field has a lot of radial MRI techniques,” Dr. Feng said. “We have many papers and substantial technical advances in this research area, but most of these techniques are not clinically used.” Despite the wide range of extensions, he saw little appetite from scanner manufacturers to adapt these for medical MRI platforms. “I started thinking, why is that?” And the question began shifting his perspective on the pursuit of better radial techniques. “I started thinking more about clinical value and impact,” he said. “Although GRASP is a clinical product, it’s mostly used for patients who are not able to hold their breath or for physicians who are specifically interested in exploring this technique.”
In other words, even though GRASP delivers the benefits of radial sampling to medical MRI, in most exams that don’t require these particular advantages, Cartesian acquisitions prevail.
“I thought I should write something to get people’s attention to start thinking what is the ultimate role of radial acquisition in clinical practice,” said Dr. Feng, adding that the idea had been with him for over a couple of years before he proposed it to the journal. While most review papers summarize the state of the field and identify the science that holds the most promise, “I’m trying to write from a relatively opposite perspective,” he said. “I’m trying to critique what I’ve been working on for a decade.”
First-line choices
In a nutshell, Dr. Feng’s critique is that the tradeoffs inherent in radial MRI have not been so decisively positive for the method to become, as he put it, a “widely-used first-line choice.”
“One question I have been thinking more and more about is: if we didn’t have this technique, would we be losing anything in clinical practice?” said Dr. Feng. “For now, I’d have to say that if we didn’t have radial acquisition, there would be some change in the overall practice of diagnostic radiology but probably not a dramatic one.”
It’s a tall bar.
In this counterfactual thought experiment the acquisition techniques that would have no better alternatives include such field-altering innovations as echo planar imaging (EPI)—a cornerstone of diffusion weighted imaging and functional MRI—and parallel imaging, which has made MRI significantly faster.
The most recent example that passes this test, Dr. Feng said, is artificial intelligence–based image reconstruction, which also substantially accelerates scans. Starting in 2016, Scientists at the Center for Advanced Imaging Innovation and Research were among the first to pioneer this approach, and also the first to demonstrate that it produces images diagnostically interchangeable with those obtained through traditional methods.
“One major reason techniques like parallel imaging, EPI, and AI-based recon have been widely adopted is that they are very robust and generalizable to a broader range of clinical exams,” said Dr. Feng. “That is very important.”
Radial sampling became a less prominent investigative theme at the Center for Advanced Imaging Innovation and Research in its second (2019-2024) and third (2024-2029) funding periods, giving way to artificial intelligence, biophysical modeling, and predictive analytics—though largely in pursuit of the same goals: to make imaging faster, more informative, and more flexible.
Dr. Feng said that the research his lab is conducting currently—including on denoising low-field MRI, reducing anatomical distortions in diffusion-weighted imaging, and teaching scanners to learn from individual patients’ imaging history—are guided by the big questions he’s asking in search of broader impact.
“I don’t have a clear answer,” Dr. Feng said. “But what I want to do is develop something for clinical practice that, if we didn’t have it, there would be a decisive difference.”
Related Stories
Diffusion-weighted imaging helps gauge risk of prostate cancer but is marred by anatomical distortions, low image quality, and long scan time. The NIH is funding a postdoc-led project at NYU Grossman School of Medicine to clear things up.
Imaging researchers at NYU Langone have used deep learning to turn noise against itself in order to improve low-field MRI. They’re after something much bigger than sharper images.
Li Feng, developer of fast MRI techniques, talks about going beyond speed, his path to academia, and the rewards of persistence.