More Than a Quick Peek
By Keith Loria
Vol. 20 No. 2 P. 12
Advanced MRI technologies speed acquisition and improve visualization.
Manufacturers are designing equipment to optimize performance with the newest technologies, and they are working hard to integrate the latest software and hardware technologies into their MRI scanners. Many of these improvements include incorporating digital technologies to replace the current MRI analog components. Daniel Sodickson, MD, PhD, a professor in the department of radiology at NYU Langone Health, says there's a dramatic revolution in imaging right now.
"We are at a really remarkable juncture in imaging, where the way we've done things for decades is starting to be supplanted by new models," Sodickson says. "One is a move from snapshots to streaming, away from the classic almost-art photography model of screening, where we keep the patient in a scanner for a long time and get the shots we need, to a much more modern fashion where we can basically press a button and the data just streams in."
This rapid multidimensional imaging is gaining traction in the industry, thanks to modern techniques.
"We can start the scan and gather different projections and views of the body continuously, even while the patient may be moving," Sodickson says. "And because of this advanced software we have at our disposal, we can then construct any view you may want and do it freezing out any motion."
Sodickson believes current visualization technologies are just the tip of the iceberg relative to what will be available in the future. Here is a look at some of the latest advanced visualization technologies for MRI.
The growing clinical demand for heart failure diagnosis and treatment as well as the increasing complexity and number of images that radiologists have to analyze are driving the need for more efficient tools. In the March issue of Nature, Bo Zhu, PhD, a researcher with the Athinoula A. Martinos Center for Biomedical Imaging at Massachusetts General Hospital, and colleagues described a new technique based on AI and machine learning, enabling clinicians to acquire higher-quality images without having to collect additional data.
"Image reconstruction is the computational process of how the raw data from the scanner [eg, MRI, CT, PET] becomes the final image for the radiologist to read. Conventionally, each step of this process is carefully crafted by engineers to reliably produce high-quality images," Zhu says. "The buzz in the field of image reconstruction is that many of these steps are being optimized or replaced entirely by machine learning algorithms—using far less expert domain knowledge—that demonstrate they can outperform conventional handcrafted approaches.
"Within a few short years, the performance of speech recognition improved dramatically due to many of the expert-engineered subsystems that modeled and processed various acoustical and language features being overhauled by a unified system composed of a conceptually simpler neural network architecture," Zhu continues. "The remarkable improvement in performance is due to the fact that we now let the data define its own representation and processing, with highly expressive neural networks that are far more flexible than conventional low-parameter models with fewer degrees of freedom."
Zhu and his research team sought to bring this paradigm shift to the expert-engineered systems of medical image reconstruction in order to push the envelope of imaging performance—in particular, enabling high-quality imaging with lower signal-to-noise (SNR) raw data, ie, faster MRI scans and lower CT and PET radiation dose. The seismic shifts in the ways that data-driven deep learning was transforming machine learning tasks such as speech recognition inspired Zhu and his colleagues to develop Automated Transform by Manifold Approximation (AUTOMAP).
"As it is trained on data derived from real scans, AUTOMAP learns to generate optimal computational strategies to process the raw scanner data, resulting in noise-robust image reconstruction for a variety of imaging protocols and modalities," Zhu says.
AUTOMAP can theoretically work well for almost any imaging scenario, but Zhu notes the most gains for scans that are SNR limited, such as fast undersampled MRI, diffusion MRI, and low-dose PET and CT.
"We also see significant imaging improvement for acquisitions that currently utilize iterative and/or model-based reconstruction techniques with assumptions that may change during the actual scan, resulting in amplification of noise and other artifacts," he says. "For example, the actual sampling trajectory in MRI for non-Cartesian spiral or radial scans may deviate from the designed trajectory that the reconstruction is based on. In these cases, we observe AUTOMAP producing more accurate image reconstructions."
Zhu says one of the differentiating aspects of AUTOMAP among other deep learning reconstruction techniques is that even the tomographic spatial domain transform (eg, Fourier and Radon), is learned. This ability to reconstruct arbitrary signal encodings allows it to extend outside of the traditional acquisition archetypes and can enable the discovery of novel pulse sequences without the need to manually devise a corresponding reconstruction.
"This capability can also computationally compensate for hardware imperfections such as field inhomogeneities or gradient nonlinearities that systematically distort the encoded signal, allowing for reduced manufacturing costs," Zhu says. "Taking this concept to the extreme, an AUTOMAP-based reconstruction system can enable entirely new and unconventional hardware designs untethered to traditional constraints and mindsets regarding spatial encoding and its reconstruction."
Beyond the image quality performance that deep learning reconstruction techniques such as AUTOMAP can provide, one common benefit of these feed-forward neural networks is that the speed of image reconstruction is nearly instantaneous—on the order of tens of milliseconds—which may not be the case for many types of medical imaging scans now. Zhu says this immediate feedback could be very helpful for clinical radiology workflow, as it enables real-time decision making during a scanning session, for example, to repeat a scan and zoom in on a particular region of interest at higher resolution.
Fabien Beckers, PhD, CEO of Arterys, which offers a medical imaging cloud platform, says there are many emerging technologies in cardiac MR visualization, from multimodality fusion to holographic visualization to its own 4D Flow.
"4D Flow changes the paradigm for cardiac MR (CMR)," he says. "It's a much faster sequence. All major [original equipment manufacturer] vendors have their own sequence, and it often requires less than 10 minutes for flow and anatomy data."
Beckers explains that Arterys started with 4D Flow CMR, which refers to phase-contrast CMR with flow encoding that is resolved relative to the three dimensions of space and the dimension of time along the cardiac cycle: 3D + time = 4D.
"Flow assessment has long been used in the evaluation of cardiovascular disease. In recent decades, the importance of improving our understanding of physiological and pathophysiological blood flow conditions has increased," Beckers says. "2D cine phase-contrast CMR is arguably the gold standard for flow volume quantification, but conventional cardiac MR is often considered a costly, lengthy, and complex exam, not to mention that it can be uncomfortable for patients."
Arterys Cardio AI takes these large studies—thousands of images—and renders them much faster than ever before using cloud computing. The user can see the full heart and measure anywhere retrospectively. In addition, Arterys has added AI to automate certain aspects of the rendering, making the CMR analysis more cost-efficient; this addresses many of the pain points of conventional CMR.
"4D Flow is a single acquisition volume that is straightforward and enables flow through any plane across it to be calculated retrospectively and with good accuracy," Beckers says. "The Arterys platform integrates with the workflow, since the images are sent directly from the scanner. The user can immediately see the automated landmarks and other automatic calculations through AI and also make any edits before publishing the report or easily share the study with colleagues."
Arterys Cardio AI is used for visualization and quantification of flow, ranging from basic aspects such as flow volume to more advanced features such as the estimation of hemodynamic effects at the vessel wall and myocardium. There are four main clinical areas in which 4D Flow is used: aortic disease, valvular disease, congenital heart disease, and shunts.
"Centers implementing 4D Flow in routine clinical practice have demonstrated that the ability to visualize comprehensive flow has allowed them to confirm or identify new diagnoses, specifically for valvular disease, in which the regurgitant jet is often eccentric, or in shunts, which might be several shunts," Beckers says. "With 4D Flow, the clinician can precisely visualize and measure the regurgitation in valvular disease and also confirm the location and flow through multiple jets."
Beckers believes this is only the beginning of what this type of reconstruction can accomplish and sees more deep learning and AI algorithms being powered by cloud computing in the future.
"The possibilities are endless," Beckers says. "Improving efficiency is key, and having a full quantitative report and advanced visualization in seconds could really improve the workflow and allow radiologists to focus on high-level diagnosis. Predictive analytics that look at historical cases and empower the physician with historical information and patient context will change how we use population and personalized health."
Mike M. Ghazal, CEO of Zetta Medical Technologies, LLC, says the company has listened to its customers and their economic concerns because not everyone can afford to acquire the newest MRI scanners.
"Zetta is working diligently to find solutions to enhance the performance of legacy scanners for these customers," Ghazal says, and this led to the idea behind Zetta's ZOOM software. "Our customers wanted to integrate a solution on their legacy MRI scanners that enhances image quality while allowing them to scan more patients."
Zetta, in addition to being a national third-party service provider, is also a medical imaging software developer; ZOOM is a fully automated solution that works in the background.
"Once the initial setup is completed, Zetta's clinical applications specialists ensure that all protocols are optimized and configured to automatically transfer images to ZOOM," Ghazal says. "Its high-performance computing engine processes the images within seconds and transfers the results to PACS systems for the radiologist to review. The automated operation is seamless to the MRI operator. They scan as usual and the ZOOM images arrive at the radiology PACS station for review without intervention."
Today, ZOOM is compatible with all MRI scanners and can be used with all types of scans, though Ghazal notes it's best to use ZOOM for long scans, such as true T1 scans.
"Scan times at four minutes or higher can be reduced to one-half the scanning time without impacting image quality," Ghazal says. "Customers looking to improve image quality can use ZOOM to enhance the capabilities of legacy scanners and hardware. ZOOM will do the work for you to maximize the performance of your existing scanner."
— Keith Loria is a freelance writer based in Oakton, Virginia.