Lab Talk

Li Feng on Rapid Imaging, Professional Trajectories, and Never Giving Up

Li Feng, developer of fast MRI techniques, talks about going beyond speed, his path to academia, and the rewards of persistence.

Li Feng, PhD, is an incoming associate professor of radiology at NYU Langone Health. He holds a doctorate in biomedical imaging and technology from NYU Grossman School of Medicine. He completed postdoctoral training at NYU Langone before moving on to research and faculty positions at Memorial Sloan Kettering Cancer Center and Icahn School of Medicine at Mount Sinai. In February 2023, he returned to NYU Langone and the Center for Advanced Imaging Innovation and Research as director of rapid imaging.

Dr. Feng is best known for the development of an MRI technique called GRASP, short for golden-angle radial sparse parallel imaging, which he led as a PhD candidate with colleagues at NYU Grossman School of Medicine and continued to extend since. GRASP has been used in more than 150,000 clinical MRI exams to-date at NYU Langone, and Siemens has implemented the method on its MRI scanners under the name Compressed Sensing GRASP-VIBE. Our conversation was edited for clarity and concision.

You have just returned to NYU Langone and the Center for Advanced Imaging Innovation and Research—this time as an independent investigator. What are some of the research directions you plan to pursue?

We are aiming to build up a program to advance rapid imaging, roughly based on the GRASP framework, which allows us to do rapid, continuous data acquisition. GRASP MRI has been primarily focused on radial sampling, but we have found this to be relatively limiting, so we want to expand the framework to Cartesian, spiral, and also EPI or propeller trajectories. The applications for these techniques include low-field MR, MR-guided radiotherapy, and diffusion imaging. That’s the broad mission of the program.

What are the potential advantages of adapting GRASP to these different types of acquisitions? 

Let me give two examples. One is that low-field MR can benefit from spiral sampling, which allows us to sample more data points with each shot than does radial sampling. Two is that the EPI trajectory is widely used in diffusion imaging, where the preparation time is long and we want to acquire as much data after each preparation as possible.

You’re one of the founders of GRASP, and you started working on it as a PhD candidate here. Since then, you have embarked on a systematic exploration of the technique. The extensions you have led include XD-GRASP, Koosh-ball GRASP, GROG-GRASP, RACER-GRASP, UTE-GRASP, GRASP-Pro, XD-GRASP-Pro, MP-GRASP, and more. What is it about this technique that keeps you interested?

Yes, I developed GRASP as a graduate student but I also want to say that it was a team effort involving many people, including Tobias Block, Hersh Chandarana, Daniel Sodickson, and Ricardo Otazo. I’ve kept working in this direction because I think the technique can be really useful. GRASP can go beyond fast imaging—it can provide a better way, or a new way, to do MRI in general. And right now, we’re trying to further optimize and develop different variants and deliver more information, more accurate information, and new information. 

Can you elaborate on what it means to do fast imaging that goes beyond speed?

When we first had parallel imaging, decades ago, the aim was to just image faster, because MR is relatively slow. Nowadays, especially with deep learning, we can achieve a lot in terms of acceleration, and I’d say it’s hard to push imaging speed much further. But MRI is still relatively slow compared to other modalities, such as CT, so we should think about different ways—and this is happening in the field with other techniques like MR fingerprinting and MR multitasking—to deliver a new framework that improves the utility of MR, extracts more information, and increases the value of MR.

GRASP is a continuous acquisition, allowing the patient to breathe normally during the scan. And right now, one area we area we’re extending GRASP acquisition to is quantitative imaging—using GRASP for free-breathing quantitative imaging while extracting more accurate information and new information that we didn’t have before. 

Multimodal imaging has also become a trend. One example is the combination of MR with focused ultrasound. Another is combining MR with the LINAC machine. We are trying to provide high-quality MRI with fast data acquisition so that those images can guide treatment. 

In these cases, we’re talking about MRI providing real-time guidance—is that right?

That’s the ultimate goal, but right now it’s challenging. We can do 2D real-time imaging but 3D real-time imaging with low imaging latency is still difficult, and that’s something we’re working on. 

You mentioned deep learning, and that’s a set of tools that has significantly shaped the field of biomedical imaging research within the last five or six years. Can you talk about how machine learning has influenced your perspective on GRASP and its potential?

Currently, the primary focus of using deep learning for MR acquisition and reconstruction is still to improve imaging speed, such as in the fastMRI project. We’re interested in using deep learning in combination with GRASP to do quantitative imaging, and I think deep learning can really be helpful here. 

Conventional quantitative imaging is a two-step approach: image reconstruction first, parameter estimation second. We can use iterative model-based reconstruction to combine the two steps but it’s computationally expensive. Here’s where deep learning has the potential to efficiently estimate parameters directly from undersampled images, and that aligns with our goal of extending GRASP beyond qualitative imaging.

Can you talk about the computational demands of GRASP reconstruction and how they’ve changed over time? 

Initially, more than a decade ago, the GRASP reconstruction took about 10 to 15 minutes per slice in MATLAB, so probably several hours for the whole MRI exam. Later, we made a lot of extensions to the reconstruction algorithm and developed a technique called GROG-GRASP, which can bypass some time-consuming steps used in the conventional setting. We also take advantage of the availability of GPUs, so right now, with the latest GRASP implementation, each slice takes about one minute to reconstruct in MATLAB.

Currently, the primary use of GRASP in the clinic is dynamic contrast-enhanced MRI, so the we don’t need the images immediately. Even if there were any flaws in the exam and we found them right away, we wouldn’t be able to repeat the scan because we cannot re-inject the contrast. So it’s okay to have the images cued for the radiologist maybe half an hour after the exam.

This fall will mark the 10th anniversary of GRASP. But we’re also in a time of round-numbered MRI anniversaries in general. Paul Lauterbur’s foundational paper on MRI just turned 50. Last year, the International Society for Magnetic Resonance in Medicine celebrated its 40th year, and the Society for Cardiovascular Magnetic Resonance marked its 30th. These are all occasions for reflection. What are your thoughts on how far the field has come and hopes for how far it can yet develop?

During the past several decades, the MR field has experienced remarkable progress, and we’re still exploring new areas. At this time, people are not just focused on MR imaging itself but are thinking about how we can combine MR with other modalities, how we can improve the value of MR, and how we can increase the accessibility of MR. These are the hot research topics we’re exploring now, and I think there’s still a lot we can do. 

Daniel Sodickson often mentions that MR is treated as an end-stage diagnostic modality. One area many people are working on now is how to bring MR to the front of the diagnostic workflow as more of a screening technique. One example here is Hyperfine’s Swoop portable scanner. I think that’s a very interesting direction to push the field toward.

Let’s talk about how you found your way into MRI research. Before you enrolled in the PhD program in biomedical imaging and technology, you were studying at the Polytechnic University (now NYU Tandon School of Engineering). How did you learn about radiology research at NYU Langone and what was your initial interest in it?

I came to New York in 2007 for a master’s in biomedical engineering at the Polytechnic University. By chance, I met a classmate who was working as an onsite Siemens scientist at NYU Langone’s radiology department, and he introduced me to the team. Daniel Kim, a cardiovascular MR expert who was at the time at NYU Langone, asked whether I’d be interested in working on a project with him. That’s how I started my career in MR.

After I got admitted to the PhD program, I asked Daniel Sodickson whether he had any funding to support me as a PhD student. Dan accepted me and asked if I’d be interested in exploring compressed-sensing image reconstruction with Ricardo Otazo. It was a very new and hot topic at that time and has turned out to be a very good choice. In 2011, we connected with Tobias Block and we were one of the first teams to work on a combination of compressed sensing and radial sampling. I found it super interesting and just stayed in this area since.

So, you entered the field of medical imaging by serendipity. What would you say to students of biomedical engineering or electrical engineering, who may not have biomedical imaging on their radar, about what this field has to offer? 

I think students in these areas have really broad choice. It’s a very tough decision between industry and academia. Biomedical imaging combines different fields—mathematics, engineering, physics, machine learning—and can be a good choice for people who have an engineering background and want to do medicine-related research and stay in academia. To me, this is a fantastic area. 

Why do you think academia has been such a good fit for you? 

It’s very simple. I want to have flexibility, and I like writing. Grant proposals, papers, these are key components in an academic career. I prefer writing a paper or a proposal to writing code. Another really important thing is flexibility: in academia, nobody requires you to start at 9:00 a.m. and get off at 5:00 p.m.

You earned your undergraduate degree in China. How did you make the transition from Sun Yat-sen University in Guangzhou to the Polytechnic University in New York?

I wanted to study abroad but didn’t have a good GPA or a good GRE score, and I was rejected by all the schools I had applied to. After that, I stayed in China and worked as an engineer for an ultrasound company. I wasn’t sure whether I could still go to the U.S., although it was something I had really wanted to do as an undergraduate. But I didn’t give up and I applied again—and got accepted to a master’s program at the Polytechnic University. That was the only admission offer I got, and it changed my life. 

I got a small stipend as a master’s student, but I still needed my parents to support the majority of my tuition and expenses. The pressure was really high, because a master’s program is very short. So from the first day in the U.S., I was thinking, what can I do after graduating? Fortunately, I was introduced to the imaging group at NYU Langone, started working with Daniel Kim, and got some salary support from him. Dan taught me a lot of hands-on skills and how to do research. I really learned a lot. He was my first supervisor and I’m so grateful. I published a paper on the work I did with him, and that’s what helped me get an offer from the biomedical imaging PhD program at NYU Grossman School of Medicine.

There actually was another PhD program I wanted to get into, at Northwestern University. I knew that my grades and GRE scores were not good enough, so I thought: how can I distinguish myself? I started writing weekly letters to Debiao Li, who was then leading cardiovascular imaging research at Northwestern—an area I was very interested in at that time. Every week, I would summarize what I’d learned. I wanted to show that I was bettering myself. I would type and print every letter and send it in a large manila envelope. 

How did this letter-writing campaign work out?

I didn’t get into Northwestern but Debiao Li offered me a research associate position for a year to gain more experience. At the same time, I got an offer from NYU and didn’t want to wait a year to start PhD studies. But we became friends. 

I was very fortunate to get to know the people at NYU Langone and get an opportunity here. As a graduate student I published a number of papers—as I said, I found that I really enjoy writing. I’m not a smart person but I work very hard. 

Are there any lessons from that time that have been especially important to you?

One of the reasons Dan Kim really liked me as a summer intern and master’s student was that I always gave him more than he expected. When he asked me for anything, I always tried to bring him more. So he was really happy with me, gave me a lot of support, wrote me a great recommendation letter, and also spoke to Dan Sodickson about my experience. That was really helpful. I pushed myself to really work hard. There were many students smarter than me, but I took a very different trajectory. Eventually, the return was really good.

Related Resources