Data-Driven Motion Correction in Clinical PET: A Joint Accomplishment of Creative Academia and Industry

TO THE EDITOR: I read with great interest the recent JNM article by Walker et al. comparing data-driven and hardware-driven motion correction technologies in PET (1). The former is an important innovation, and its transition into the marketplace is exciting to see. Publications such as this one play a pivotal role in the technology’s acceptance and broader dissemination. However, this work is very similar to work from our group published in 2016 (2), and unfortunately, our publication was not properly referenced.

Like Walker et al., we compared nongated, software-gated, and hardware-gated images head-to-head in a large set of clinical PET scans, using quantitative analysis of lesion uptake and qualitative masked reviewer scoring of image quality, with similar results—a statistically significant preference for software-gated images over hardware-gated images and with similar ratios of performance metrics. There are, of course, subtle differences between the gating approaches, and Walker et al. note that their work validates newly available commercial technology. Given that this work focused on commercial product testing, it should add scientific context to note that the key points they presented also describe our earlier findings.

Also, in their closing discussion Walker et al. suggest that data-driven gating with quiescent-period sorting is a practical motion correction strategy but that retention of more than 50% of coincidences may be required before respiratory gated PET imaging can dependably support the clinic. We are happy to share that we have also studied this issue, finding that clinical PET data support a spectrum of ideal or optimal bin sizes throughout a given population and, ultimately, that no single bin size will ensure maximum benefit, or even any benefit, for any given patient (3). The implication, and what we have shown in our work, is that the legacy of one-size-fits-all binning strategies could be improved upon with a data-conforming binsize one, and make the motion correction effort better suited for routine clinical use.

The commercial technology discussed in the article of Walker et al. is GE Healthcare’s MotionFree product. To the credit of the company, it recognized the potential of data-driven motion correction and developed a product to translate this potential to clinic settings. The algorithms used in GE Healthcare’s product, and in our 2016 and earlier publications (2–5), are remarkably similar.

Data-driven motion correction has evolved over the last two decades, and our group has been active in its development. In 2007, we recognized that, at the data level, motion in PET is captured and recorded in localized signal fluctuations. To our knowledge, we were the first to demonstrate the ability to characterize patient motion through direct constructive combination of time–activity signal fluctuations in the data acquired, an original idea that at the time improved significantly on the strategy of tracking geometric or center-of-mass–type motion (4–7). In recognizing the importance of practicality, our group was also the first, to our knowledge, to consider and demonstrate that processing can be accelerated to virtually real time through strategic collapsing of raw (i.e., sinogram) data (8). Notably, these innovations provided proof of principle and formed the basis of most data-driven gating publications since. Additionally, we believe that our group was the first to discuss and demonstrate the concept of fully automated workflows as a uniquely practical strategy for bringing robust motion management into the clinic (9–11). We developed innovative spinoff concepts, such as using a quality factor (defined as the ratio of signal in respiratory and nonrespiratory temporal frequencies in our collected motion trace) to determine a priori the capacity of the signal to usefully correct a patient scan (7) and to modulate bed acquisition times based on information from such signals for practical clinical integration (10). It is gratifying that the MotionFree product integrates all the foregoing innovations originally presented in our earlier publications.

The overlap between our motion characterization innovation and the principal-component analysis algorithm supporting the GE product has not yet been articulated in literature, and is presented here for context and comparison. In the years 2007–2010, our group developed the idea of strategically combining the time evolution of raw PET signal to characterize patient motion and suggested that it is likely the methods could be improved with further development of signal weighting (5,7,8). In 2011, for example, Thielemans et al. investigated this possibility by integrating a well-established mathematic function of principal-component analysis to calculate these weighting factors (12). Our recent comparisons between principal-component analysis–based weighting and our original constructive combination-based methods have not yet been published, but they show that the 2 methods perform comparably or, in many cases, virtually identically (13)—a likely consequence of the fact they are derived from the same deconstruction of signal. It is, therefore, no surprise that the results of Walker et al.’s clinical assessment and ours are so similar. This is an important result because it indicates that the data-driven gating technology, based on combining spatially clustered signal fluctuations, can perform comparably across different centers, vendors, and implementations.

In data-driven motion management, our field is witnessing the culmination of a physics innovation concept-to-impact cycle, with GE Healthcare providing a first-to-market product (for general PET respiratory motion correction). Many research scientists who began this journey over a decade ago have contributed original ideas to this effort (12,14–20). Alongside others, our group contributed to inventing the technology, enabling its practicality, advocating for its consideration, and demonstrating its clinical utility. In the process, we found researchers eager to cooperate, vendors who offered support, and an effective process for solution development that built off each other’s accomplishments and ideas. We also found challenges, which illuminated prospects to expand our field’s infrastructure to better support data-driven innovation. These opportunities include evolving our understanding of data as a resource; opening pathways for data innovation to reach the market/clinic; and fostering a community that embraces new concepts for innovation, which we expect to come with a rapidly advancing digital landscape (21–23).

Ultimately, our goal should be to transition to a field where data science innovation is only limited by our imagination and not by a legacy infrastructure, and we are presented now with a chance to build that field. The path there is best supported with allied cooperation, inclusive visions, and shared successes.

• © 2021 by the Society of Nuclear Medicine and Molecular Imaging.