Impact of Acquisition Time and Misregistration with CT on Data-Driven Gated PET

Abstract

Objective:

Data-driven gating (DDG) can address patient motion issues and enhance PET quantification but suffers from increased image noise from utilization of <100% of PET data. Misregistration between DDG-PET and CT may also occur, altering the potential benefits of gating. Here, the effects of PET acquisition time and CT misregistration were assessed with a combined DDG-PET/DDG-CT technique.

Approach:

In the primary PET bed with lesions of interest and likely respiratory motion effects, PET acquisition time was extended to 12 min and a low-dose cine CT was acquired to enable DDG-CT. Retrospective reconstructions were created for both non-gated (NG) and DDG-PET using 30 sec to 12 min of PET data. Both the standard helical CT and DDG-CT were used for attenuation correction of DDG-PET data. SUV_max, SUV_peak, and CNR were compared for 45 lesions in the liver and lung from 27 cases.

Main Results:

For both NG-PET (p=0.0041) and DDG-PET (p=0.0028), only the 30 sec acquisition time showed clear SUV_max bias relative to the 3 min clinical standard. SUV_peak showed no bias at any change in acquisition time. DDG-PET alone increased SUV_max by 15±20% (p<0.0001), then was increased further by an additional 15±29% (p=0.0007) with DDG-PET/CT. Both 3 min and 6 min DDG-PET had lesion CNR statistically equivalent to 3 min NG-PET, but then increased at 12 min by 28±48% (p=0.0022). DDG-PET/CT at 6 min had comparable counts to 3 min NG-PET, but significantly increased CNR by 39±46% (p<0.0001).

Significance:

50% counts DDG-PET did not lead to inaccurate or biased SUV – increased SUV resulted from gating. Improved registration from DDG-CT was equally as important as motion correction with DDG-PET for increasing SUV in DDG-PET/CT. Lesion detectability could be significantly improved when DDG-PET used equivalent counts to NG-PET, but only when combined with DDG-CT in DDG-PET/CT.

INTRODUCTION

Respiratory motion and misregistration are both well-established issues in positron emission tomography/computed tomography (PET/CT) imaging. PET and/or CT gating are the traditional solutions for addressing problems with respiratory motion. Options for PET gating generally fall into three strategies: one based on the use of an external device (EDG) (Nehmeh et al., 2004; Werner et al., 2009; Pan et al., 2004; Hamill et al., 2020b; Meier et al., 2018), and two others that utilize the newer method called data-driven gating (DDG). EDG requires setting up the necessary hardware and making modifications to the standard clinical workflow. As a result, EDG is fully prospective, requiring appropriate screening and selection of patients in advance who may benefit from PET gating. While achievable, this makes clinical workflows more cumbersome and raises concerns whether all relevant patients should be assigned for gating. Historically, the only option for gating of PET and/or CT was with EDG, which complicated applications enough that it was rarely used outside of radiation therapy planning (Pepin et al., 2014).

In DDG, raw PET data is analyzed to extract motion information and accomplish gating without relying on an external device (Buther et al., 2016; Buther et al., 2020; Walker et al., 2020; Kesner et al., 2014; Schleyer et al., 2009; Kesner et al., 2018; Liu et al., 2010). The first strategy for DDG-PET involves retrospective application of the DDG process to any patient of interest. In principle, all patients are eligible since the gating information is derived directly from raw data and no other hardware is needed. In the second application of DDG-PET, patients are not screened beforehand but rather prospectively monitored for its potential application. This is often achieved by identifying relevant motion for each PET bed and using thresholds to decide if gating with DDG may be beneficial (Kang et al., 2021; Messerli et al., 2021; Sigfridsson et al., 2021). Then certain PET beds are chosen for DDG and the acquisition time is increased only for those portions of the PET exam. The increased acquisition time offsets the fact that nearly all DDG-PET methods utilize less than 100% of total PET counts. With the prospective monitoring strategy, a patient may have zero, one, or even up to four PET beds (2.5 min/bed for non-gated PET, 5 min/bed for DDG-PET) where DDG is applied on a high sensitivity (22 cps/kBq) PET/CT scanner (Messerli et al., 2021).

DDG-PET has recently seen an increase in clinical use and applications, likely due to the availability of commercial options (Khamis and Wollenweber, 2019; Feng et al., 2020; Hong et al., 2014). Importantly, DDG-PET has been shown recently to be either equivalent to (Buther et al., 2016; Buther et al., 2020; Kvernby et al., 2021) or preferred over EDG PET (Walker et al., 2020). A key advantage of gating with DDG is motion correction that reduces lesion volumes, generally leading to increased standard uptake values (SUV) and more accurate representations of uptake (Sigfridsson et al., 2021; Walker et al., 2020; Buther et al., 2016; 2013). The degree of change in SUV from DDG varies in previous studies, but most have shown at least an average increase of 10% (Buther et al., 2016; 2013; Kang et al., 2021) with some as high as 25% (Sigfridsson et al., 2021). In cases with significant respiratory motion combined with PET/CT misregistration, SUV can be increased as much as ~65% on average (~500% for specific lesions) once these issues are corrected with DDG-PET and DDG-CT (Pan et al., 2020a; Thomas and Pan, 2021). An additional benefit of DDG-PET previously discussed is enhanced lesion detectability (Walker et al., 2020; Messerli et al., 2021). Small, low uptake lesions located in regions with significant motion may be “blurred out” and missed with standard PET.

PET image noise is increased in most DDG-PET methods because they preserve data at the end-expiration (EE) respiratory phase with minimal motion and discard data between end-inspiration (EI) and EE with significant motion. There have been several different DDG-PET techniques explored previously, but most utilize a process along these lines and retain ≤50% of the total PET data (Buther et al., 2016; Khamis and Wollenweber, 2019; Thielemans et al., 2011; Kesner et al., 2014). This leads to one of two situations: 1) keep the same acquisition time relative to non-gated (NG) PET and accept increased noise in DDG-PET images (Buther et al., 2016; Pan et al., 2020a), or 2) increase acquisition time to counteract the loss in DDG-PET counts and maintain image noise comparable to NG-PET (Walker et al., 2020; Kang et al., 2021; Sigfridsson et al., 2021; Messerli et al., 2021). If the more prospective patient monitoring strategy is used for DDG-PET, increased acquisition times may alleviate concerns about reduced counts from the DDG process. However, the total scan time is then unknown beforehand (Messerli et al., 2021), potentially making patient scheduling more difficult. The effect that the more prospective DDG-PET protocols may have on clinical workflows and patient scheduling is likely different for each clinic – busier clinics may struggle to easily adopt such strategies.

DDG-PET methods have also been explored where most or all PET data is used. In these techniques, motion parameters are also derived from the raw PET data and non-rigid registration or motion deblurring methods are applied to combine separate phases into a single PET dataset (Lu et al., 2018; Chan et al., 2018; Buther et al., 2020). While it is true that these methods likely avoid the problem of increased noise, they may introduce other issues related to the use of non-rigid registration, interpolation of counts, and lack of matched attenuation correction (AC) for specific PET phases. At present, there is still debate regarding whether the potential benefits of DDG-PET outweigh the necessary drawbacks related to its widespread adoption (Pan et al., 2020a; Buther et al., 2016; Kesner et al., 2014). With the use of higher sensitivity PET scanners with a larger detector coverage of 25 to 30 cm, the trade-off between image noise and acquisition time adjustments may become less of an issue.

When using DDG only for PET, registration between DDG-PET and CT is left to chance in the commonly used free-breathing CT acquisition. When PET data and the CT used for AC are not matched with respect to respiratory phase, significant changes in SUV can occur and the accuracy of PET/CT registration is compromised (Meier et al., 2019; Pan et al., 2020a; Thomas and Pan, 2021). In fact, these changes were shown to be even more significant in gated PET relative to NG-PET (Meier et al., 2019; Pan et al., 2020a; Thomas and Pan, 2021; Messerli et al., 2021). Clinical applications for PET/CT imaging rely on accurate SUV values and sufficient image quality for diagnostic purposes. As a result, misregistration between DDG-PET and CT is an issue that warrants full attention and must be addressed. Some DDG-PET protocols in use also recognize the importance of matching CT with DDG-PET by attempting to acquire a breath hold CT at the EE respiratory phase using patient breathing instructions (Kang et al., 2021; Sigfridsson et al., 2021).

The tradeoffs in the clinical application of DDG-PET make it important to understand the impact that relevant factors have on its performance. There has been ample previous work related to the effects of acquisition time and/or administered activity on NG-PET (Boellaard et al., 2015; Carney et al., 2004; Akamatsu et al., 2015; Yan et al., 2016; Lodge et al., 2012; Namias and Jeraj, 2019). Previous studies have also addressed the importance of registration between PET and CT, as well as potential solutions (Pan et al., 2005; Chi et al., 2007; Chi et al., 2008; Hamill et al., 2020a; Pan et al., (in press)). With regard to DDG-PET, fewer relevant studies are available. Most have only assessed a single DDG-PET acquisition time and only had a standard helical CT available for AC. As DDG-PET continues to grow in clinical use, more studies are beginning to explore whether DDG-PET acquisition time should be increased so that DDG-PET counts match with standard, NG-PET (Walker et al., 2020; Kang et al., 2021; Sigfridsson et al., 2021; Messerli et al., 2021). Along the same line, the potential impact of gated PET’s increased image noise on SUV accuracy, overall image quality, and lesion detectability have been explored in select cases (Buther et al., 2016; Meier et al., 2018; Walker et al., 2020). But these studies were limited to comparisons between gated PET and NG-PET at only one level of equivalent total counts (or mismatched total counts). At present, it remains unclear whether reduced counts DDG-PET produces false increases in SUV that are labeled as motion correction effects. But it is also unknown to what extent increased noise from reduced counts degrades real effects from motion correction on PET quantification. Specifically with respect to changes in lesion detectability from gating with DDG-PET, the potential for DDG to not only increase SUV but also increase image noise must be better understood (Kesner et al., 2014).

In this work, our primary goal is to assess the impact of acquisition time and misregistration with CT on DDG-PET with objective, quantitative analysis. Clinically relevant metrics analyzed include lesion maximum SUV (SUV_max), peak SUV (SUV_peak), and contrast-to-noise ratio (CNR). Background signal-to-noise ratio (SNR) was also assessed for impressions on general image quality. We have evaluated these metrics for both NG-PET and DDG-PET across a wide range of acquisition times that includes both increased and decreased times relative to common clinical standards. Finally, we analyzed the influence of misregistration between DDG-PET and CT on the clinical metrics tracked by utilizing both helical CT and DDG-CT for AC of DDG-PET. Our results show that acquisition time impacts the various quantitative aspects of PET/CT imaging in different ways. As a result, the optimal strategies for DDG-PET may depend on the specific clinical application. We also conclude that accurate registration between DDG-PET and CT is crucial to optimize its benefits for PET/CT imaging applications.

METHODS

Patient Selection

The study was approved by an Institutional Review Board. The need for informed written consent was waived. The patients and cases for this study came from a previous clinical trial at our institution that occurred from 2014–2016, with the data analyzed and presented here collected retrospectively. The details of this clinical trial have been discussed previously (Meier et al., 2019). Briefly, 37 patients who were referred for PET/CT were prospectively recruited. Recruitment was based on patients with lesions between 1 and 3 cm in diameter in locations typically affected by respiratory motion (lower thorax and upper abdomen). Three patients were excluded from analysis in the present study due to missing or corrupted data no longer available relative to the prior study referenced above. Three further patients were excluded on the basis of respiratory waveforms for accurate gating, as discussed further below. Finally, four patients were not included for any analysis related to lesions due to prospective lesions of interest being either not visible or non-avid. Overall, there were 31 cases available for the study, with 27 utilized specifically for lesionrelated analysis.

PET/CT Acquisition Protocol

The scans were all acquired with a GE Discovery 690 PET/CT scanner with 15.4 cm axial and 70 cm transaxial fields of view. The injected activity was targeted at 370 MBq of ¹⁸F-fluorodeoxyglucose (FDG). The mean ± σ (1 standard deviation) uptake time prior to imaging was 64.8 ± 7.0 min (Meier et al., 2019), in line with the current protocol for FDG-PET/CT imaging at our institution. A free breathing helical CT scan was acquired first with the following scan parameters: 120 kVp, 60–560 mA tube current (tube current modulation, TCM, was used), a noise index of 30, 1.375 pitch, 0.5 s rotation time, and a 64×0.625 mm (4 cm) beam width. Whole-body PET was acquired next. The PET axial field of view (“bed”) containing the lesion(s) of interest utilized a 12 min acquisition time, a 400% increase over the clinical standard of 3 min used at that time (hereafter referred to as the primary bed). All PET beds outside of the primary bed had 3 min acquisition times except for the legs, which were 1.5 min.

After the PET acquisition was complete, a cine CT was also acquired, but only over the primary PET bed region. The step and shoot cine CT protocol utilized 120 kVp, a constant tube current of 10 mA (no TCM), a 0.5 s rotation time, a scan coverage of 8×2.5 mm (2 cm) per segment, and a cine duration for each segment equal to the patient’s full breathing cycle time plus one second. Note that some details of this cine CT protocol were different from the more recent DDG-PET/CT protocol in use at our institution (Pan et al., 2020a; Thomas and Pan, 2021; Pan et al., (in press)). The estimated effective radiation dose from the current version of the protocol is 1.3 ± 0.6 mSv (Pan et al., 2020a). Effective doses for cases in this study were likely a bit lower due to reduced mA and rotation time settings.

DDG-PET and DDG-CT

In the previous trial, the PET/CT cases used in this study had gating achieved for both PET and CT using EDG (Meier et al., 2019). In this study, the listmode PET data and raw cine CT data were retrospectively analyzed to produce DDG-PET and DDG-CT. All retrospective PET images were reconstructed with time-of-flight ordered subsets expectation maximization, including point-spread function modeling (VPFX-S). Reconstructions utilized 24 subsets and 2 iterations, with a 6.4 mm Gaussian post-filter. Note that the specific PET reconstruction settings chosen were based on the current clinical settings at our institution. These settings derive from a balanced discussion between physicist recommendations for optimization and physician preference for image characteristics. The vendor recommends an additional iteration for PSF-based reconstruction but those settings were not used in this work.

The commercially available GE product called Q.Static was used for all DDG-PET analyses without any modifications. DDG-PET data were derived from a portion of the total PET data (target: ~50%) in the EE (quiescent) phase at 30% offset from the EI phase of each respiratory cycle. DDG was applied to the primary PET bed, as well as the two adjacent PET beds on opposite sides due to the 15% overlap between beds in acquisition. GE’s Q.Static assigns a motion magnitude parameter (R value) to all PET beds analyzed for DDG, and the user can decide on an R value threshold to set for applying gating (Khamis and Wollenweber, 2019). The default R value threshold for applying Q.Static is 15 (Khamis and Wollenweber, 2019). We used this default value as a guide to identify the most appropriate cases for gating in this study.

There were four cases with an R value <15 in the primary PET bed where maximal respiratory motion was expected, with specific R values of 7.0, 10.0, 11.8, and 13.0. After additional review of the waveforms from these cases by a medical physicist, all except for the one with an R value of 7.0 were deemed to be too inconsistent for appropriate gating with DDG. Despite the low R value of 7.0, the remaining case’s waveform was well-tracked and consistent so we included it in the study. The 31 PET datasets in this study had R values with mean ± σ of 22.0 ± 10.0 and a range from 7.0 to 67.0 in the primary PET bed. Six cases had R values between 15 and 20 while 24 had R > 20. For the PET beds adjacent to the primary bed, DDG was applied if the R value for that PET bed was above the threshold used for the primary bed (e.g. a threshold of 15 for all cases except for one, with R = 7.0). Some adjacent PET beds had very low R values so we determined the relevant threshold for a case on the basis of the primary PET bed rather than adjacent beds.

The cine CT data was used to derive an EE CT, or DDG-CT, which replaced the cine CT region in the baseline helical CT. The derivation of DDG-CT from the cine CT data has been outlined previously (Pan et al., (in press)) and will not be repeated here. Three different PET/CT combinations were then possible: 1) NG-PET with helical CT AC (NG-PET), DDG-PET with helical CT AC (DDG-PET), and 3) DDG-PET with DDG-CT AC (DDG-PET/CT).

PET Simulated Acquisition Times

Both NG-PET and DDG-PET retrospective reconstructions were created to simulate various acquisition times. On the clinical workstation, the only option for creating reduced counts reconstructions while also employing DDG with Q.Static is to use specific fractions of the full data – e.g. for a 3 min reconstruction only the first 3 min of data can be used. It is not possible to randomly sample the data using the clinical workstation. The clinically acquired 12 min NG-PET served as the starting PET dataset, and then the first 6 min, 3 min, 1.5 min, 45 sec, and 30 sec of listmode data were used to create simulated NG-PET reconstructions. Similarly, the 12 min NG-PET listmode data was used to create 12 min, 6 min, 3 min, 1.5 min, 45 sec, and 30 sec DDG-PET reconstructions using Q.Static. Note that 30 sec is the minimal acquisition time needed to utilize DDG with GE’s Q.Static. Because the DDG process only utilized a portion of the total PET data, the effective acquisition times for DDG-PET were all reduced. For consistency, however, we will refer to all NG and DDG-PET data based on their clinical acquisition times throughout this study unless specifically addressed otherwise. Overall, 15 different PET/CT datasets were created: both NG-PET and DDG-PET at 12 min, 6 min, 3 min, 1.5 min, 45 sec, and 30 sec; and DDG-PET/CT at 12 min, 6 min, and 3 min only. The effects of DDG-CT were sufficiently observed without assessing DDG-PET/CT at all acquisition times.

Quantitative Measurements

Only lesions that were within the primary PET bed with a 12 min clinical acquisition time were included. Based on the design of the PET/CT acquisition protocol, all lesions analyzed were therefore also within the cine CT region. This produced a total of 45 lesions in the 27 cases analyzed, with 18 in the liver and 27 in the lung. The largest number of lesions a single case contributed to the total was four. The lesions were contoured with a threshold method based on SUV_max. An initial threshold of 40% was used where possible, but for some lesions it was adjusted upwards until differentiation from background was achieved. Once the threshold percentage was established for a particular lesion, it was used consistently across all PET/CT datasets compared. Of the 45 lesions analyzed, 29 used a 40% threshold, 7 used a 50% threshold, 6 used a 60% threshold, and 3 used a threshold >60% with the maximum threshold being 75% (n=1). Because SUV values changed with both acquisition time and the CT used for AC, unique lesion contours were created for the range of different PET/CT datasets analyzed. While PET segmentation methods are varied and worthy of comparison generally (Hatt et al., 2017), this work aimed to use a simple, objective contouring method so that quantitative measurements of lesions could be compared fairly among all PET/CT datasets.

SUV_max and SUV_peak (Hyun et al., 2016) were calculated for all lesions and all PET/CT datasets. SNR was determined for background (BG) regions in the lung and liver. For BG SNR, a cylindrical region of interest (ROI) ~10 cm³ in size was placed in an area with consistent uptake that was also completely within the primary PET bed. Careful attention was paid to avoid any regions with variable uptake and also ensure the ROI would be consistently located in well-registered soft tissue areas where applicable (e.g. liver) for all relevant PET/CT datasets analyzed. The above restrictions limited the selection to only 23 liver and 13 lung BG regions, with some cases having both liver and lung while others were liver or lung only. In one case, neither BG region could be placed adequately. Once established for a particular case, the same exact BG ROIs were used consistently for all PET/CT datasets analyzed. SNR was defined as follows:

$$BGSNR=\frac{BGROISU{V}_{mean}}{BGROISU{V}_{\sigma }}$$

where SUV_σ represents the standard deviation of SUV values.

Lesion CNR ratios were also calculated, defined as follows:

$$LesionCNR=\frac{LesionSU{V}_{max}-LocalBGSU{V}_{mean}}{LocalBGSU{V}_{\sigma }}\text{,}$$

where Local BG was established as the local BG region near each individual lesion. Local regions near lesions often maintained different SUV characteristics relative to the overall BG. Additionally, the use of different CT datasets for AC impacted registration between PET and CT. As a result, local BG regions could change substantially between different PET/CT datasets. Some lesions were misregistered between PET and CT in either NG-PET or DDG PET, both of which used the helical CT. The most common issue was a lesion being localized in air in CT but uptake in PET, leading to incorrect AC. As a result, SUV values both inside lesions and in the regions surrounding lesions were impacted. So Local BG SUV_mean and SUV_σ values were obtained for each individual lesion by placing circular ROIs (~1.2 cm²) near the lesion at each slice, in the most reasonable area of consistent uptake. Lesion CNR was calculated using SUV_max instead of SUV_mean in part because SUV_max should not be dependent on contouring method.

Ratios for all quantitative metrics of interest were calculated for all other PET/CT datasets relative to NG-PET (AC = helical CT) with a common acquisition time of 3 min as follows:

$$MetricRatio=\frac{MetricinModifiedPET/CT}{Metricin3\text{min}NGPET}$$

In some cases, other PET/CT datasets were used as a reference instead when more specific changes were of interest. Potential distinctions between liver and lung were assessed for both lesion and BG metrics in the most clinically relevant PET datasets of 3 and 6 min DDG-PET and DDG-PET/CT.

Statistical and Bland-Altman Analysis

We identified nearly all datasets analyzed in this work (44 of 45) as lognormal with both Anderson-Darling and Shapiro-Wilk normality tests after applying the natural logarithm. As a result, parametric significance tests were deemed appropriate after taking the natural logarithm of all relevant data. Mean ± σ was used for comparisons. For all matched comparisons among paired datasets (e.g. effects of acquisition time, DDG, or CT for attenuation correction), repeated measures one-way ANOVA analysis was used. If a statistically significant difference among all groups was established, multiple comparisons were assessed for more specific analysis among groups. A false discovery rate correction using the two-state step-up method of Benjamini et al.(Benjamini et al., 2006) was also implemented using a false discovery rate of 0.01, with all p values reported after adjustment for multiplicity. For comparisons among diverse groups, (e.g. liver vs. lung), unpaired t-tests with Welch’s correction were used. All statistical analyses were performed with GraphPad Prism 9.2.0 (GraphPad Software, San Diego, California, USA) and statistical significance was considered true for p < 0.01.

In order to better understand the combined effects of motion correction with DDG-PET, improved AC with DDG-CT, and changes in noise with reduced counts inherent to DDG-PET, Bland-Altman style analysis (Bland and Altman, 1986) was also employed. Percentage changes in lesion SUV were calculated as follows:

$$\Delta SUV=\frac{SU{V}_{new}-SU{V}_{ref}}{SU{V}_{ref}}\times 100.$$

The reference dataset for SUV_ref differed depending on the comparison of interest. Mean percentage changes and 95% prediction intervals (±1.96σ, PI) of the distribution of changes were also calculated as outlined in Bland-Altman analysis (Bland and Altman, 1986). Limits of agreement (LoA) were determined as mean ± 95% PI. The maximum and minimum LoA for expected changes in SUV under specific PET/CT conditions provided baselines that established limits for intrinsic changes with ~95% statistical confidence. Changes in SUV outside of the LoA could then be identified as unique, enabling the separation of different contributions from noise, motion correction, and improved registration and AC.

RESULTS

SUV_max Bias and Acquisition Time

Figure 1 shows mean ± σ SUV_max ratios for NG-PET and DDG-PET where three different acquisition times were used as the reference dataset. Figure 1(a) presents the effects of reduced acquisition time on SUV_max bias across the widest possible range available in this study by using 12 min as the reference dataset. In NG-PET, only the 30 sec acquisition time, with highly decreased counts relative to 12 min, showed statistically significant SUV_max bias (mean: +7%). For DDG-PET, a significant bias was observed at both the 45 sec acquisition time (mean: +8%) and the 30 sec acquisition time (mean: +15%). As shown in Fig. 1(b), for ratios relative to the 3 min acquisition time, no bias in SUV_max was observed for NG-PET or DDG-PET until the significantly reduced 30 sec acquisition time (NG-PET: +6%, DDG-PET: +13%). Increasing acquisition time above 3 min also did not bias SUV_max. Finally, when the 30 sec acquisition time served as the reference, all other acquisition times showed significant (negative) bias for both NG-PET and DDG-PET.

Figure 1.

Plots of SUV_max ratios for NG-PET and DDG-PET at all acquisition times using three different reference datasets for the ratios: (a) 12 min, (b) 3 min, and (c) 30 sec. The mean values are plotted and the errorbars represent ±σ. All data is plotted with acquisition time decreasing from left to right. For those data showing a statistically significant difference to the reference dataset, the level of significance is indicated with * p<0.01 and ** p<0.001.

One clear observation from Fig. 1 was the difference in SUV_max bias and variability between NG-PET and DDG-PET at equal clinical acquisition times. This mainly stems from the fact that DDG-PET uses approximately 50% counts and therefore has an effective acquisition time that is reduced relative to NG-PET. Figure S1 in the Supplementary Information shows a different way of calculating and plotting the data from Fig. 1(b), where effective acquisition time was used instead of actual acquisition time. In this case, it was clear that both NG-PET and DDG-PET maintained similar SUV_max bias and variability when on equal footing with approximately equivalent acquisition time and counts.

Lesion SUV_max and SUV_peak

Table 1 presents detailed summary data of lesion SUV_max, SUV_peak, and CNR ratios for all the PET/CT data analyzed in this work. The ratios were calculated relative to 3 min NG-PET for all NG and DDG-PET acquisition times. Ratios calculated relative to 3 min DDG-PET for all DDG-PET acquisition times are also included, as well as the statistical significance of all of the comparisons tested. Figure S2 in the Supplementary Information presents boxplots of the all the same ratios from Table 1 calculated relative to 3 min NG-PET so that the full distributions of data can be seen and compared if desired.

Table 1.

Lesion SUV_max, SUV_peak, and CNR summaries for all NG and DDG-PET

	Ratio Relative to 3 Min NG-PET &
	SUVmax		SUVpeak		Lesion CNR
NG-PET	Mean ± σ	p	Mean ± σ	p	Mean ± σ	p
12 min	1.00±0.07	0.20	1.01±0.05	0.097	1.24±0.36	***
6 min	1.00±0.04	0.18	1.01±0.03	0.10	1.21±0.31	***
1.5 min	1.00±0.08	0.20	1.00±0.05	0.26	0.84±0.29	0.0006
45 sec	1.03±0.12	0.12	1.01±0.07	0.23	0.79±0.31	***
30 sec	1.06±0.13	0.0041	1.02±0.08	0.12	0.72±0.33	***
DDG-PET
12 min (5.9 min)	1.13±0.19	***	1.08±0.12	***	1.28±0.48	0.0022
6 min (2.9 min)	1.15±0.20	***	1.08±0.11	***	1.17±0.39	0.028
3 min (1.4 min)	1.16±0.21	***	1.07±0.11	***	0.97±0.40	0.026
1.5 min (41 sec)	1.16±0.18	***	1.07±0.10	***	0.92±0.43	0.0036
45 sec (19 sec)	1.22±0.23	***	1.08±0.12	0.0001	0.77±0.37	***
30 sec (11 sec)	1.28±0.27	***	1.09±0.14	***	0.72±0.40	***
	Ratio Relative to 3 Min DDG-PET #
DDG-PET	SUVmax		SUVpeak		Lesion CNR
12 min (5.9 min)	0.99±0.10	0.17	1.01±0.04	0.20	1.37±0.44	***
6 min (2.9 min)	1.00±0.07	0.50	1.01±0.03	0.067	1.25±0.32	***
1.5 min (41 sec)	1.01±0.08	0.47	1.00±0.04	0.28	1.00±0.44	0.022
45 sec (19 sec)	1.06±0.18	0.050	1.01±0.08	0.49	0.84±0.41	***
30 sec (11 sec)	1.13±0.23	0.0028	1.02±0.12	0.31	0.77±0.41	***

Abbreviations: σ = standard deviation

^&All ratios and p values are relative to 3 min NG-PET

^#All ratios and p values are relative to 3 min DDG-PET

^***p<0.0001

As outlined in Fig. 1(b) and Table 1, SUV_max was biased by +6% on average when acquisition time was reduced from 3 min down to 30 sec in NG-PET. On the other hand, Table 1 and Fig. S2(b) show that SUV_peak was much more robust against changes in noise with acquisition time. The variability across lesions was much lower and even at 30 sec there was no statistically significant bias. When using either 12 min or 30 sec as the reference dataset (like Fig. 1(a,c) for SUV_max), no statistically significant bias was observed at any acquisition time for SUV_peak. For DDG-PET, the gating process led to increased SUV_max relative to 3 min NG-PET for all acquisition times. The same was true for SUV_peak, though the average increases were smaller in magnitude and less variable across different acquisition times. Just like with NG-PET, there were no statistically significant biases in SUV_peak at any acquisition time for DDG-PET, when referenced to 12 min, 3 min, or 30 sec acquisition times.

The effect of DDG-CT on SUV_max ratios relative to NG-PET can also be seen in Fig. S2(a) with the data for DDG-PET/CT. Summary data for these SUV_max ratios are included in Table 2. When registration was improved with DDG-CT, DDG-PET/CT increased SUV_max relative to NG-PET significantly more than DDG-PET alone. At all acquisition times analyzed (3, 6, and 12 min), DDG-PET/CT maintained higher SUV_max than DDG-PET. The typical mean SUV_max ratio for the overall change from NG-PET to DDG-PET/CT was ~1.33. Roughly half of the increase in SUV_max for DDG-PET/CT came from the gating process with DDG – the typical mean SUV_max ratio for DDG-PET relative to NG-PET was ~1.15 (Table 1). The other half originated from optimized AC with DDG-CT – the typical mean SUV_max ratio for DDG-PET/CT relative to DDG-PET was also ~1.15 (Table 2).

Table 2.

Lesion SUV_max, SUV_peak, and CNR ratio data for DDG-PET/CT

	Ratio Relative to 3 Min NG-PET &
	SUVmax		SUVpeak		Lesion CNR
DDG-PET/CT	Mean ± σ	p	Mean ± σ	p	Mean ± σ	p
12 min (6 min)	1.31±0.44	***	1.24±0.33	***	1.54±0.65	***
6 min (3 min)	1.33±0.42	***	1.24±0.32	***	1.39±0.46	***
3 min (1.5 min)	1.33±0.42	***	1.24±0.32	***	1.11±0.39	0.13
	Ratio Relative to 3 Min DDG-PET #
DDG-PET/CT	SUVmax		SUVpeak		Lesion CNR
12 min (6 min)	1.14±0.33	0.011	1.16±0.30	0.0004	1.68±0.69	***
6 min (3 min)	1.15±0.29	0.0011	1.16±0.29	0.0004	1.52±0.49	***
3 min (1.5 min)	1.15±0.28	0.0011	1.15±0.29	0.0004	1.18±0.27	***

Abbreviations: σ = standard deviation

^&All ratios and p values are relative to 3 min NG-PET

^#All ratios and p values are relative to 3 min DDG-PET

^***p<0.0001

The results for SUV_peak were similar, but also distinct in some ways. The typical mean SUV_peak ratio for the overall change from NG-PET to DDG-PET/CT was 1.24, smaller than that for SUV_max. This could result from the ability of SUV_peak to minimize the effect of noise. But the typical mean SUV_peak ratio for DDG-PET/CT relative to DDG-PET was also very similar to SUV_max at ~1.16. The exact quantitative impact of motion correction on SUV was more difficult to pin down due to the effect of noise from reduced counts. But for both SUV_max and SUV_peak, improved AC with DDG-CT increased SUV values by 15–16% on average.

Figure 2 shows Bland-Altman style plots of changes in SUV_max and SUV_peak. Figure 2(a) plots the percentage change in SUV_max relative to the actual SUV_max, with 3 min DDG-PET compared to 6 min DDG-PET as the reference. This data provides a sense of the intrinsic variability and potential bias associated with 50% reduced counts in DDG-PET. Figure 2(d) shows the same type of plot for SUV_peak. There was no clear bias in SUV_max from reducing acquisition time by 50%, with a mean ΔSUV_max of ~0.5%. LoA ranged from −12.6% to +13.5% due to the inherent variability from increased noise. These LoA values provide a baseline for the random changes in SUV_max from simply reducing counts by 50%.

Figure 2.

Bland-Altman style plots for changes in SUV_max (a-c) and SUV_peak (d-f). Parts (a,d) compare 3 min DDG-PET relative to 6 min DDG-PET, (b,e) compare 3 min DDG-PET relative to 3 min NG-PET, and (c,f) compare 3 min DDG-PET/CT relative to 3 min NG-PET. The mean change and limits of agreement (LoA) for 3 min DDG-PET relative to 6 min DDG-PET are included as dotted and solid lines, respectively, in each set of figures. The LoA values are also included in (a,d). In (b,c,e,f) the lesions are identified by their change in SUV_max or SUV_peak relative to the LoA from (a,d): black circle = within the LoA, blue diamond = above the LoA, and red square = below the LoA. The percentages of lesions falling in each group relative to the LoA are included in each figure. The mean changes are plotted as dash-dot lines in (b,c,e,f) for reference.

In Fig. 2(b), 3 min DDG-PET is compared to 3 min NG-PET as the reference. Roughly 60% of lesions had changes in SUV_max still within or below (negative) the expectations for noise contributions alone. But about 40% (n=19) of lesions showed increased SUV_max that went beyond the maximum LoA. These results indicate the increased SUV_max for these 19 lesions was very likely not solely due to noise – motion correction with DDG-PET also played a key role. In Figure 2(c), the effects of DDG-CT on SUV_max changes can be readily seen. Many lesions showed further increases in SUV_max, and an additional eight lesions had SUV_max changes beyond the maximum LoA. This moved the total to well over half of all lesions. While it must be acknowledged that noise may contribute to increases in SUV_max for DDG-PET at 50% counts, the results in Fig. 2 provide strong evidence that motion correction is primarily responsible for many lesions. This is especially true when DDG-PET is matched with DDG-CT for optimal AC.

The data for SUV_peak in Fig. 2 help to support these conclusions further. SUV_peak was more robust against changes in noise from reduced counts, having a mean ΔSUV_peak of −0.7% and smaller LoA range from −5.5% to +4.1%. With DDG-PET alone, a similar number of lesions showed increases in SUV_peak beyond the maximum LoA. But then with DDG-PET/CT in Fig. 2(f), more lesions had further increases in SUV_peak as compared to SUV_max. This led to >75% of all lesions above the maximum LoA for SUV_peak with DDG-PET/CT. When intrinsic changes in SUV due to noise alone were more effectively handled with SUV_peak, the impacts of motion correction with DDG-PET and optimal AC with DDG-CT were even more visible.

Lesion CNR

Figure S2(c) and Table 1 show that for NG-PET, increased acquisition times led to increased CNR and decreased acquisition times produced decreases in CNR. But changes in lesion CNR were more complicated with DDG-PET. DDG-PET at 3 min showed slightly reduced CNR relative to NG-PET at 3 min, but this difference was not statistically significant. When increasing acquisition time to 6 min so that DDG-PET had approximately equivalent counts to NG-PET, the average lesion CNR was higher but this difference was also not statistically significant. Notably, the same was true when comparing 3 min DDG-PET to 1.5 min NG-PET (p = 0.039), where counts were also approximately the same. A further increase in acquisition time to 12 min was necessary to clearly increase lesion CNR for DDG-PET above 3 min NG-PET.

Compared to NG-PET at 3 min, lesion CNR for DDG-PET/CT at 3 min was increased on average, but not statistically different. However, it was significantly increased for DDG-PET/CT at 6 min. Both of these results are different from what was observed with DDG-PET alone. As seen in Fig. 3(b), not all lesions had improved CNR with DDG-PET/CT relative to DDG-PET alone, but the average improvement was 18%. When acquisition time was increased so that DDG-PET/CT had approximately equivalent counts to 3 min NG-PET (Fig. 3(c)), lesion CNR was increased by 39% on average over 3 min NG-PET, and 29% on average over 3 min DDG-PET/CT. DDG-PET at 6 min had 15 (33%) lesions with lower CNR than 3 min NG-PET, and 3 min DDG-PET/CT had 21 (47%). DDG-PET/CT at 6 min reduced lesions with CNR worse than 3 min NG-PET to only 5 (11%).

Figure 3.

Correlation plots of (a) SUV_max ratios and (b,c) CNR ratios comparing DDG-PET/CT to DDG-PET: (a,b) compare 3 min DDG-PET/CT to 3 min DDG-PET, (c) compares 6 min DDG-PET/CT to 3 min DDG-PET. In each plot, all ratios were calculated relative to 3 min NG-PET (same data as in Tables 1 and 2). The mean ± σ percentage changes in the ratios being compared are included in each figure.

Background SNR

Figure 4 shows plots of BG SNR ratios as a function of change in acquisition time for both NG-PET and DDG-PET, and both liver (n=23) and lung (n=13). The ratios were calculated relative to 3 min NG-PET. Liver SNR mean ± σ was 16.1 ± 3.7 for 3 min NG-PET and 11.6 ± 2.7 for 3 min DDG-PET. Liver SNR changed roughly as expected with duration for NG-PET, though it was not affected quite as much as the theoretical prediction. But all NG-PET acquisition times showed statistically significant changes in liver SNR relative to 3 min NG-PET (all p<0.0001), with increased SNR for longer acquisition times and decreased SNR for reduced acquisition times. For DDG-PET, liver SNR changes tracked the predictions with theory more closely. When comparing NG-PET to DDG-PET at approximately equivalent counts, both the 6 min and 3 min groups (effective acquisition time) were not statistically different, but both the 1.5 min and 45 sec groups were (p ≤ 0.0083). All DDG-PET maintained statistically significant changes in liver SNR relative to 3 min NG-PET except for 6 min DDG-PET (p=0.044), which had roughly equivalent counts to 3 min NG-PET.

Figure 4.

Plots of background SNR ratios as a function of mean duration ratios for (a) NG-PET liver, (b) DDG-PET liver, (c) NG-PET lung, and (d) DDG-PET lung. The SNR and duration ratios are calculated relative to 3 min NG-PET – acquisition times increase from 30 sec to 12 min with increasing duration ratio on the x-axis. The theoretical prediction for SNR vs scan acquisition time is also plotted for reference in each figure. Note the different x-axis scale for DDG-PET in order to maintain clarity in the figures.

Lung SNR mean ± σ was 9.5 ± 2.1 for 3 min NG-PET and 7.7 ± 1.5 for 3 min DDG-PET. For both NG-PET and DDG-PET, lung SNR changed less with acquisition time than liver SNR. Reduced duration did not produce decreases in SNR as significant as theory predicts, and increased duration actually led to much smaller enhancements in SNR relative to theoretical predictions. Nevertheless, all changes in lung SNR were statistically significant relative to 3 min NG-PET (p ≤ 0.001), except for 6 min DDG-PET (p=0.19), just like with liver SNR. When comparing NG-PET to DDG-PET at roughly equivalent counts, no statistically significant differences in lung BG SNR were observed.

Liver vs. Lung: All Metrics

When comparing changes in SUV and CNR between liver and lung lesion groups, no statistically significant distinctions were observed. For DDG-PET, increases in SUV were nearly identical between liver and lung lesions. For DDG-PET/CT, SUV increases were higher on average for liver lesions but these differences were not statistically significant. Similar relationships were observed for changes in CNR, with liver lesions generally having higher average increases but not distinct enough to establish statistical significance. The only clear distinctions between liver and lung were related to BG SNR as already shown in Fig. 4.

Example Patient Images

Figure 5 shows example PET/CT fusion images for both liver and lung cases from this study. In Fig. 5(a), two liver lesions near the diaphragm were very poorly registered with baseline helical CT. Despite the misregistration, the effect of motion correction with DDG-PET is still visible for both lesions in the middle panel. But SUV_max values were significantly more impacted with improved AC in DDG-PET/CT (right panel). Figure 5(b) shows another liver case with a primary lesion near the liver boundary. In this case, motion correction with DDG-PET increased SUV more than improved AC with DDG-CT. This case did not have the same clear misregistration with the helical CT as the case in Fig. 5(a), but DDG-CT still increased SUV_max by ~9%. In the lung case in Fig. 5(c), motion correction played a minimal role. However, the motion correction that was achieved managed to make registration with the helical CT even worse than with NG-PET. Once registration was improved with DDG-PET/CT, SUV_max increased by 31%. The lung case in Fig. 5(d) involved a much smaller lung lesion, yet the effects of improved AC were still readily visible. Motion correction with DDG-PET increased SUV_max by 8% but DDG-PET/CT increased SUV_max an additional 24% beyond motion correction alone.

Figure 5.

Example cases showing PET/CT fusions for both (a,b) liver and (c,d) lung lesions; (a) shows axial images while (b-d) show coronal images. Left panel: 3 min NG-PET, center panel: 3 min DDG-PET, right panel: 3 min DDG-PET/CT. The SUV_max values for each lesion are labeled in all parts of each figure. The DDG-CT images from the fusions in each right panel were post-processed with PixelShine to reduce noise (Pan et al., 2020b).

Whole body, coronal PET images for the two liver cases from Figs. 5(a,b) are displayed in Fig. 6. The impact of motion correction with DDG-PET was evident for both cases, while the improved registration and AC were especially visible for the case in Fig. 6(a). All of the DDG-PET image sets (ii, iv in both cases) with 50% reduced counts showed slight increases in noise that can be discerned in the liver, leading to reduced SNR. However, these differences in overall image appearance were minimal relative to the two “full counts” image sets in 3 min NG-PET (i) and 6 min DDG-PET (iii).

Figure 6.

Whole body, coronal PET images for two different cases with liver lesions (A,B). For each case, four different PET datasets are shown: (i) 3 min NG-PET, (ii) 3 min DDG-PET, (iii) 6 min DDG-PET, and (iv) 3 min DDG-PET/CT. The primary lesions are marked with red arrows and SUV_max values for those lesions are indicated in each subfigure. Liver SNR values are included for each image as well. Case (A) is the same as shown in Fig. 6(a) and Case (B) is the same as shown in Fig. 6(b).

DISCUSSION

This study analyzed the impact of gating with DDG, PET acquisition times, and improved AC with DDG-CT in PET/CT. DDG-PET increased SUV_max relative to NG-PET – this has also been established in many previous DDG-PET and EDG-PET studies (Buther et al., 2016; Walker et al., 2020; Kang et al., 2021; Sigfridsson et al., 2021; Messerli et al., 2021; Meier et al., 2019; Meier et al., 2018). The amount of average SUV_max increase observed in this study with DDG-PET alone (~15%) was similar to previous work. But when PET/CT registration was improved with DDG-PET/CT, the average increases observed here (~33%) were more than nearly all prior studies. Even with DDG-PET/CT, the cases in this work did not yield increases in SUV as substantial as our previous work, most likely due to the biased selection of misregistered cases in those studies (Pan et al., 2020a; Thomas and Pan, 2021).

The majority of our results in this study derived from comparisons to a reference PET/CT dataset with typical clinical imaging conditions and sensible intrinsic noise equivalent counts: NG-PET with a 3 min acquisition time. Under these conditions, we determined that the use of ~50% of total PET counts in DDG-PET applications cannot reasonably cause SUV to change from image noise alone. The observed increases in SUV in both this work and previous DDG-PET studies were very likely due to gating with DDG. No clear, statistically significant bias in DDG-PET SUV_max was observed until counts were reduced to ~17% of the reference value at 3 min DDG-PET. The same results were obtained for NG-PET SUV_max. However, when the 30 sec acquisition time was used as the reference dataset, it produced biased SUV_max values relative to all other acquisition times. This was true for both NG and DDG-PET. So under imaging conditions with low intrinsic counts, even small changes in acquisition time may lead to biased SUV values.

This study only analyzed one radiotracer (FDG) with very well-established PET imaging protocols that yield sufficient counts for quality images. The specific effects of acquisition time on SUV bias may be different with other radiotracers or imaging conditions – more consistent parameters for such comparisons may be noise equivalent counts and specific uptake biodistributions. Overall, this study provided strong evidence that clinically applicable increases or decreases in acquisition time relative to the reference standard of ~3 min are not expected to produce relevant changes in DDG-PET SUV quantification. Whether DDG-PET uses 50% counts, or acquisition time is adjusted to produce equal or even 100% increased counts relative to 3 min NG-PET, SUV quantification is not clearly impacted.

Notably, in previous work a statistically significant change in SUV_max, with a roughly 6% increase over the reference full counts NG-PET, was observed at only 35% counts (Buther et al., 2016). Additionally, a different previous study that analyzed image noise on SUV_max bias in NG-PET showed that when reducing acquisition time to ~17% (like the 30 sec acquisition time in this study relative to 3 min), the expected SUV_max bias was roughly +15% (Lodge et al., 2012). Both of these previous studies conveyed results different from those observed in this work. A possible explanation is the use of more advanced iterative reconstruction techniques in this study, including time-of-flight and point-spread function modeling. It is possible that the effects of reduced acquisition time on image noise and SUV_max bias are less significant when these enhanced features are used (Adams et al., 2010; Reddy et al., 2018). In a very recent study of total body PET utilizing both time-of-flight and point-spread function modeling in iterative reconstruction, lesion SUV_max bias was also not observed until acquisition time was decreased significantly (~7% of reference time) (Hu et al., 2021).

It is important and particularly relevant for the clinical use of DDG-PET to understand how SUV_max is affected by acquisition time and reduced counts. While 50% counts DDG-PET may not suffer from consistent bias, the increased noise does lead to variability in SUV_max that must be acknowledged. Our results from Fig. 2 showed that such variability is approximately ±13% to a 95% confidence level. When analyzing specific lesions, apparent increases in SUV_max from reduced counts DDG-PET may not truly be from motion correction. On the other hand, the full magnitude of increases from motion correction may be difficult to determine if random noise decreases SUV_max. Our results for SUV_peak provided strong evidence that it is a much more stable option for assessing changes in uptake when the effects of noise are difficult to fully separate. This has been confirmed before and is one of the main designs of the SUV_peak metric (Hyun et al., 2016).

SUV_peak was not biased at any level of reduced or increased counts studied in this work, regardless of the acquisition time used as the reference dataset. It also maintained a much smaller range of intrinsic changes when counts were reduced by 50% as compared to SUV_max (Fig. 2). As a result, it was more straightforward to extract the true impacts from motion correction with DDG-PET and improved AC with DDG-CT on SUV quantification. The results from SUV_peak helped to further confirm that increased SUV from motion correction is real and not due to bias from increased noise. These results also showed that, on average, improved registration with DDG-CT was at least equally as important as motion correction with gating for optimizing SUV.

The expectation is that gating from DDG could increase lesion CNR due to motion correction and a more robust collection of counts into a smaller volume. This is especially true when DDG-PET has approximately equivalent counts to NG-PET, as the local BG noise should then be comparable. But the results for DDG-PET with helical CT for AC were underwhelming in this regard. A key aspect relates to registration between DDG-PET and CT – if registration is not optimized, the full impact of gating may not be realized and SUV values could be affected (Meier et al., 2019; Pan et al., 2020a; Thomas and Pan, 2021). This was clearly observed with our results here as well. Generally, DDG-PET/CT increased SUV_max relative to NG-PET by ~33%. Approximately half of this increase resulted from the gating process with DDG and half was due solely to improved AC with DDG-CT. Therefore, a metric like lesion CNR that relies heavily on maximizing SUV was also strongly impacted by DDG-CT. It is worth noting though that changes in lesion CNR were more complicated than changes in SUV_max, even though they are related based on the definition of CNR is this study. As seen in Fig. 3, some lesions with a positive or minimal change in SUV_max from improved AC with DDG-CT actually showed decreased CNR. Changes in AC from DDG-CT also affect the regions near lesions and as a result, lesion detectability.

The importance of appropriate registration between DDG-PET and CT cannot be understated – assessing the performance and clinical utility of DDG-PET without ensuring accurate and optimized registration will likely lead to incorrect conclusions. Had we only analyzed DDG-PET with helical CT AC in this study, logical conclusions would have been as follows: 1) DDG-PET only increases SUV_max by ~15% on average, 2) on average, lesion CNR may be degraded when using DDG-PET with ~50% of total counts, and 3) only when PET acquisition time is increased enough that DDG-PET utilizes more counts than NG-PET will it clearly improve lesion detectability (CNR). Such conclusions would put the clinical utility of DDG-PET in serious jeopardy. Instead, our logical conclusions were that DDG-PET combined with DDG-CT increased SUV_max by >30% on average, DDG-PET/CT using only 50% counts (which enables retrospective gating with DDG for all patients) maintained lesion CNR relative to NG-PET, with increased lesion CNR on average, and DDG-PET/CT with equivalent counts to NG-PET increased lesion CNR by ~40% on average. These conclusions put DDG-PET on much more solid footing and open it up for a variety of applications depending on the specific needs of the clinic.

DDG-PET with a 50% reduction in counts led to reduced BG SNR, but this was expected since a 50% reduction in counts also decreased BG SNR for NG-PET. Modern PET iterative reconstruction is not a strictly linear process, so the assumption of SNR changing proportionally with $\sqrt{time}$ is likely not applicable (Chang et al., 2012). The results for both NG-PET and DDG-PET seemed to match with this idea in that BG SNR values did not change as much with acquisition time as theory predicts. The diversion in BG SNR from the theoretical model was much more visible in the lung. This may be because the lung is a more diverse organ with both blood vessels and air, while the liver has more uniform tissue. The standard deviation for background uptake in the lung may depend on more than just acquisition time (counts).

This study did have limitations that should be discussed as well. Obtaining clinical PET data where the acquisition time is increased more than 100% over the clinical standard is difficult. We would have preferred increased patient and/or lesion numbers as well as a more balanced spread between liver and lung lesions. The lack of many clear distinctions between liver and lung results in this study may have been related to the limited number of cases and the bias toward lung lesions. Additionally, only FDG cases were analyzed in this work. Recent studies have shown that DDG-PET results may depend on the specific radiotracer used (Kvernby et al., 2021; Messerli et al., 2021). There are other radiotracers used clinically for PET – understanding their specific impact on DDG-PET is also important. However, by limiting our analysis to FDG we were able to remove the possible effect of radiotracer on the results.

The retrospective reconstructions acquired in this study may have also been limited. As described in Methods, in order to utilize Q.Static on the clinical workstation, only the first x min of the full 12 min of data can be used to simulate shorter acquisitions, where x = 6, 3, 1.5, etc. If patient breathing cycles or patterns suffered from inconsistencies throughout the full PET acquisition, the simulated reconstructions may have been affected. We attempted to minimize this effect by removing cases with clear issues on the basis of analysis of the DDG respiratory waveforms. Related to this issue is the very short acquisition time reconstructions that were part of this study. It is possible that the relative convergence of the PET reconstructions changed at these very short acquisition times with very noisy data.

An argument could be made that the cases in this study are not representative of routine PET/CT imaging, and we would agree. But we feel that the patient group in this study was well-suited for a detailed analysis of DDG-PET/CT for two reasons: 1) it is likely a reasonable representation of cases that would be chosen for DDG-PET in applications where patients are monitored for its use with motion tracking and thresholds – recall that all but one case in this study had R values >15 in at least one PET bed, and 2) there was no information prior to the PET/CT acquisitions regarding whether or not misregistration between PET and CT may occur. These conditions are different from our previous work on DDG-PET/CT (Pan et al., 2020a; Thomas and Pan, 2021). As a result, the analysis in this work should be reflective of the potential impact DDG-PET and DDG-PET/CT can have on a wider range of clinical PET/CT applications.

Another key limitation was related to the cine CT protocol used for the cases in this study. This was an older protocol that is different from the one currently used at our institution. Specifically, in the new protocol TCM is used for the cine CT as well, with a tube current limit of 20 mA, the rotation time is increased to 0.8 s, and a fixed cine duration of 5 s is used for all patients (Pan et al., 2020a; Thomas and Pan, 2021). For some of the cases involving larger patients in this study, the cine CT exposure was not sufficient. This led to negatively biased Hounsfield unit values in the DDG-CT, which in turn affected SUV quantification for those cases. It is not possible to correct this problem retrospectively, however it should be noted that some cases and lesions likely had inaccurately low SUV (and likely CNR) values for DDG-PET/CT data where DDG-CT was used for AC. For calculating lesion CNR, local BG ROIs relied on subjective placement by a human observer. But most analyses in this work were tied to relative comparisons of the same PET data analyzed at different acquisition times and/or with different CT datasets for AC – it was not necessary that the quantitative metrics compared be correct, only consistent. A subjective reader study to assess the impact of acquisition time and DDG-CT on lesion detectability and image quality would be a prudent future avenue of research. Finally, DDG-PET as used in this study relies on motion analysis only from the PET acquisition, and in many current clinical applications its prospective use is determined from motion-based thresholds derived from PET data. In our experience, we have observed some cases where motion was not identified readily in PET but it was from the cine CT shortly after the PET acquisition. The DDG-CT used in this work for AC of DDG-PET can also be used to identify end-inspiration (EI) phase CT images. As a result, it is possible to estimate the respiratory motion magnitude by comparing EI and EE phase CT images (Pan et al., (in press)). Future work from our group will focus on exploring this method in more detail and comparing the DDG-CT motion analysis with the motion results extracted from DDG-PET.

The overall results from this study show that there is not a single clinical workflow or optimal use of DDG-PET that is the clear choice for all applications. If DDG-PET is to be used primarily to optimize SUV and ensure the accuracy of diagnostic/treatment response metrics tied to SUV quantification, our results suggest that acquisition time is not critical. If DDG-PET is applied retrospectively to any case of interest, where a ~50% reduction in counts cannot be avoided, relevant biases in SUV are highly unlikely on the basis of a standard acquisition time of ~3 min. As a result, prospective monitoring of patient motion and increased acquisition times to match counts with NG-PET are likely not necessary just to ensure non-biased SUV. Under these conditions, it is more difficult to extract with 100% certainty the true impact of motion correction on SUV_max. The use of SUV_peak for this purpose is likely a more robust alternative.

If the primary concern is lesion detectability and ensuring PET offers the best possible chance for accurate diagnosis, DDG-PET may need to be used differently. Retrospective application of DDG-PET to all patients, where DDG-PET suffers from reduced counts, can likely only maintain lesion CNR relative to NG-PET, not necessarily increase it. And the results from this study showed that this is only true for DDG-PET/CT, not DDG-PET alone. If registration between DDG-PET and CT is left to chance as in standard applications of DDG-PET, our results indicated that total PET acquisition time must be increased more than 100% over NG-PET to ensure lesion detectability is improved on average. This is likely not practical at most PET/CT clinics. DDG-PET workflows where patients are prospectively monitored and PET acquisition time is doubled to offset reduced counts can enhance lesion detectability – but DDG-CT is likely necessary to make this occur consistently. The issue of reduced counts DDG-PET and the question of changing acquisition times and altering clinical workflows for gated PET beds may be addressed by full counts DDG-PET methods that manage to use all of the acquired PET data (Chan et al., 2018; Buther et al., 2020). Our group is currently working on full counts DDG-PET methods as well, and this is an excellent topic for future work to compare to the results from this study and others associated with 50% counts DDG-PET.

A final aspect of this discussion relates to the body regions of interest for correcting respiratory motion. Generally speaking, upper abdomen (liver, etc.) is different from lower lung. Lesion detectability is likely less of a concern for the lung since background uptake is typically minimal and lung lesions can often be identified with CT alone. Baseline lesion CNR values in the lung are often very high, especially relative to the upper abdomen (Meier et al., 2018). Therefore, improvements in CNR may not be as relevant. Similarly, if maintaining overall image quality in DDG-PET is a primary goal for certain applications, this is likely more readily achieved in the lung.

The situation is likely different in the upper abdomen, however. Background uptake of FDG is often high enough in the liver and other organs in this region that contrast and image quality both become relevant concerns. For other radiotracers such as those based on ⁶⁸Ga that target neuroendocrine lesions, such issues may be even more concerning (Kroiss et al., 2013; Kuyumcu et al., 2013). Reduced count DDG-PET may very well lead to lower diagnostic quality images under these conditions. In a recent study where DDG-PET was found to be generally improved over EDG PET, it was cases with smaller, low uptake lesions in the upper abdomen where DDG was preferentially chosen (Walker et al., 2020). These types of cases are where prospective monitoring of patient motion, increased acquisition times to offset reduced counts, and ensuring DDG-PET and CT are well registered all become more relevant. In the end, each institution must prioritize its goals with respect to DDG-PET and design workflows that help to achieve these goals. We hope that our work in this study helps to provide guidance for these decisions as the use of DDG-PET continues to grow clinically.

CONCLUSION

DDG-PET using ~50% reduced counts did not lead to inaccurate or biased SUV when derived from a standard acquisition time of ~3 min for NG-PET. Instead, DDG-PET/CT increased SUV_max by >30% and SUV_peak by ~25%, on average. Misregistration with CT impacted lesion SUV significantly, with DDG-CT AC increasing SUV at least an equivalent amount to the gating process with DDG-PET. Changes in lesion detectability and overall image quality were more directly tied to acquisition time. DDG-PET with approximately equivalent counts to NG-PET maintained comparable image quality but did not lead to significantly improved lesion CNR. When combining DDG-PET with DDG-CT in DDG-PET/CT, lesion detectability was dramatically improved over NG-PET with equivalent counts. Overall, the effects of acquisition time and misregistration with CT on DDG-PET must be carefully considered depending on its specific clinical application. Clear benefits are achievable with DDG-PET/CT, but these must be weighed along with its potential drawbacks and impacts on clinical workflows.