Positron emission tomography (PET) imaging combined with [¹⁸F]2-fluoro-2-deoxy-D-glucose ([¹⁸F]FDG) is a widely used tool in research settings and clinical diagnostics (Fletcher et al., 2008; Gupta et al., 2011; Rischka et al., 2018; Hahn et al., 2020). The most common semi-quantitative metric that can be derived from static PET acquisitions is the standard uptake value (SUV), which is used as a surrogate of glucose metabolism for quantifying FDG uptake. Nevertheless, significant limitations of the accuracy of SUV remain. One of these limitations stems from the fact that SUV cannot provide a reliable measure of the tissue kinetics and does not account for the tracer levels in plasma (Huang, 2000; Boellaard, 2011).

Dynamic PET imaging provides spatio-temporal metabolic characteristics when combined with kinetic modelling methods and may exhibit greater robustness than simplistic SUV measures (Patlak et al., 1983; Wu et al., 2001; Cheebsumon et al., 2011). A full kinetic analysis comprises the solving of differential equations and on the voxel-level variance can be high due to increased noise which leads to less reliable parameter estimates and thus an inconsistent kinetic model. Therefore, graphical methods such as Patlak analyses are being widely used in both clinical and research settings due to its robustness and simplicity. This is particularly feasible for [¹⁸F]FDG as the radiotracer exhibits almost irreversible kinetics within a common scan time of less than 60 min. Furthermore, studies have shown that glucose metabolism estimated with the Patlak plot represents a more accurate and robust index of glucose metabolic rate when compared to SUV (Freedman et al., 2003).

Another disadvantage of modelling procedures is the requirement of invasive blood sampling to produce an accurate arterial plasma input function (AIF). Image derived input functions (IDIFs) represent a promising non-invasive alternative, aiming to extract the AIF from a suitably large region containing a robust blood pool in PET images and may be supplemented with venous blood to improve accuracy to the IDIF (Hahn et al., 2012, 2013). To date, whole body (WB) PET scans were acquired using a multibed, multipass protocol whereby each bed position requires scan times of 2–5 min (Boellaard et al., 2015). This however decreases temporal resolution greatly and thus affects the accuracy of kinetic modelling. The introduction of WB continuous bed motion (CBM) protocols utilizing state-of-the-art PET/CT systems offer a promising solution to these issues. Here, dynamic PET data can be acquired with increased spatio-temporal resolution while still keeping the kinetic information intact (Osborne et al., 2015; Osborne and Acuff, 2016). This is achieved by combining a short dynamic scan over the cardiac region to automatically acquire the IDIF and subsequent multiple WB sweeps (van Sluis et al., 2021). Combined with a high temporal resolution at the start of the protocol this further mitigates spillover effects. This then disseminates to improved estimation of the input function and outcome parameters (Viswanath et al., 2020) and thus better quantitative images (Wang et al., 2020; Zhang et al., 2020; Dimitrakopoulou-Strauss et al., 2021). Another option used in brain activation studies would be using a reference region to normalize the PET image; this however is only semi-quantitative, thus limiting accuracy (Clark et al., 1985; Dukart et al., 2010; Hua et al., 2015).

In theory, such an approach would represent a clinically feasible protocol, which provides quantitative estimates of glucose metabolism and high subject throughput.

In this work, we aim to validate the automatically generated IDIFs derived from the aorta and left ventricle using a Biograph Vision Edge (Siemens Healthineers, Germany) PET/CT system and included vendor-specific software with its recommended settings to the reference standard AIF (https://www.siemens-healthineers.com/molecular-imaging/options-and-upgrades/software-applications/flowmotion-multiparametricpet-suite). Furthermore, IDIFs were also extracted manually and images were corrected for motion. Finally, brain glucose metabolism was computed with the Patlak plot using the automatic vendor-based approach as well as manually generated IDIFs and AIFs.

Although the validity of the IDIF, using various blood pools and tracers, has already been assessed (Van der Weerdt et al., 2001; Hahn et al., 2012, 2013; Sari et al., 2021), the comparison of a vendor specific, fully automated IDIF software has yet to be evaluated. As PET/CT scanners become more advanced and image quality improves, automatic quantification software solutions will become more commonplace. Therefore, it is paramount to assess the strengths and limitations of this type of solution. If the vendor specific software were proven to accurately match the AIF, it would greatly improve accessibility to molecular imaging research and reduce labor needed in a clinical setting. This comparison study aims to provide arguments to what extent the fully automated IDIF software is suitable for absolute quantification in both clinical and research settings. We used the cerebral metabolic rate of glucose (CMRGlu) and WB net influx constant Ki to assess the effects each input function has on tissue quantification.

Out of the seventeen participants recruited, one was dropped due to a failure in arterial sampling. Furthermore, the vendor-based automatic generation of all three VOIs was successful for only 15 participants. Visual inspection of the automatic vendor generated VOIs showed a suboptimal placement of both the aorta and LV regions. The center of the aorta VOI was frequently located in the aortic wall, whereas the LV was seen over the ventricle wall. Furthermore, the peak value in the automatically generated aorta IDIF was on average lower than that of the AIF and manually placed VOIs. The LV had higher peak values than AIF, but not statistically significantly, (p _LV/blood = 0.45 and p _aorta/blood = 0.27, Figure 2A), while the AUC of both scanner generated IDIFs were lower than the AIF (Figure 2B). Finally, the LV IDIF peaked on average 15 s ± 5 s earlier than the AIF. Similarly, the aorta IDIF peaked on average 12 s ± 6 s earlier to the AIF, see Figure 2C. Supplementary Table S1 shows the coefficient of variance for organs of interest.

The purpose of this study was to assess the validity of a fully automatic IDIF and the resulting quantitative values of CMRGlu and WB Ki using a WB CBM [¹⁸F]FDG PET/CT protocol to the reference standard of AIF. We observed that the vendor generated VOIs were not always optimally placed, resulting in suboptimal input function extraction. The left ventricle VOI was more susceptible to motion artifacts stemming not only from its cardiac function but also body movement. However, once corrected from above mentioned artifacts, it showed the best agreement to the AIF (Figure 3). Nevertheless, both quantitative values derived from IDIFs and the IDIFs themselves indicated a good match to the AIF. Furthermore, a simple solution to correct for motion which was applied after the scan resolves uncertainties in IDIF generation, which may represent an important benefit for long scan protocols and patient groups more prone to movement (Figures 3, 4; Table 2).

Firstly, the similarity between the automatically vendor generated IDIFs, namely, (aorta and left ventricle) were compared to the AIF. Overall, we observed a good agreement in both AUC and the peak values from all three sources. Both IDIFs displayed an earlier peak onset when compared to the AIF, which most probably emerges from anatomical distance between the heart and the radial artery. Furthermore, we observed that the IDIFs derived from the LV had a higher variance than those derived from the aorta. This can be attributed to the motion intense physiological left ventricular function of the heart as well as potential spillover effects from the adjacent myocardium. Peak values derived from the scanner-generated aorta IDIF were generally lower than both the LV and AIF but more stable than the LV. The differences between the automatic pipeline when compared to the other 3 could be in part attributed to the automatic misplacement of both the aorta and LV VOIs. However, this was not the case for the manually derived aorta IDIF. Previous studies suggest that the aorta is a more robust option for the IDIF (Sari et al., 2021), which is also supported by our observations (Table 1c) as it is less susceptible to movement artifacts. Visual inspection of the automatic vendor generated VOI placement revealed sub-optimal placement for both the aorta and LV, which influences not only IDIF extraction but also the quantitative values generated. Similar peak and onset values extracted from the aorta and LV IDIFs were also found when using the new high-performance uEXPLORER scanner (Zhang et al., 2020). Zhang et al. also show that cardiac motion in the LV affects the TAC extraction, in their analysis the aorta was selected as the TAC extraction was not as affected by motion. They further compare smaller VOIs like the carotid, brachial and femoral arteries and show that these VOIs are affected by dispersion and partial-volume effects due to their size (Zhang et al., 2020).

Subject motion during scans can severely affect the placement of regions used for IDIF extraction, which in turn will affect modelling outcomes. For graphical analysis using the Patlak plot an incorrectly defined IDIF will affect the area under the curve and thus directly bias CMRGlu values (Yao et al., 2021). This can be seen in our results where we compared the CMRGlu values derived from both the manually and automatically defined IDIFs. Therefore, it is imperative to either choose a VOI less prone to motion artifacts or correct for such motion. One option to overcome this is to use a population derived input function, which by definition is not affected by motion (Rahmim et al., 2019; Naganawa et al., 2020; van Sluis et al., 2021). This method however is also not optimal as the input function is only scaled by the participants’ blood which must be extracted venously or arterially during the scan (Brock et al., 2005; Vriens et al., 2009). This can lead to bias (Zanotti-Fregonara et al., 2012; Naganawa et al., 2020) and may not account for individual as well as pathological variations in the shape of the input function. The low correlations between the LV and AIF AUC before movement correction are indicative of either misplaced VOIs or movement during the scan. This can also be seen when analyzing the IDIFs of individual subjects. We found manually placing the VOIs for each individual considerably improved the robustness of the IDIFs, as we were able to adjust for each individuals varying anatomy. By further correcting for movement artifacts, the LV IDIF displayed a better match to the AIF than the aorta. However, it is acknowledged that this may not be feasible in clinical settings. Similarly, using rigid-body transformations to correct for motion in the areas where the IDIF was extracted, the agreement between the IDIF and AIF further improved. Thus, the combination of manually placed VOIs and motion correction yielded the best results.

Since the peak values and AUC of the aorta IDIF were on average lower than that of the AIF, the CMRGlu was higher for both the scanner-generated and manually drawn aorta IDIF when compared to the AIF. Nevertheless, a high correlation between all three CMRGlu estimates was found (De Geus-Oei et al., 2006; Naganawa et al., 2020) and especially CMRGlu estimates for the grey matter were in line with previous literature (Liu et al., 2021; Sari et al., 2021). Of note, there were only a few exceptions to this, showing unphysiologically high Ki values. In most cases this can be corrected by reprocessing the data, which would require a trained technician to check each step of processing. This is time consuming and not feasible in many clinical scenarios. These problems did not appear when using manually drawn IDIFs or the AIF.

While dynamic PET imaging yields not only a more accurate clinical picture but also more spatio-temporal metabolic characteristics i.e., metabolic rate and distribution volume. This however comes with a cost, comprising prolonged scan time, complex protocols, longer image reconstruction times and invasive arterial blood sampling, which might not be tolerable for certain patient populations e.g., dementia patients. Therefore, static PET imaging utilizing SUV as a surrogate for metabolic rates has been used to reduce the workflow requirements of complex dynamic PET imaging. SUV is affected by several factors, including patient habitus, blood glucose levels and uptake time. All of these factors greatly reduce the information available for diagnosis (Braune et al., 2019).

We would like to acknowledge certain limitations of our work. Even though the AIF can be used as an input to the vendor pipeline, this data must be acquired from an external third party device. Thus, the comparison the AIF as an input for the automatically generated Ki images was not analyzed, consequently limiting the interpretability of the intrinsic, vendor specific quantification algorithm. Furthermore, as the vendor software comes as is, it was only possible to partly recreate a similar extraction and quantification pipeline, thus also limiting comparability between not only the IDIF’s but also the Ki image generation. In these instances, the vendor software was substituted with PMOD’s kinetic modelling software. Future development and validation of abovementioned approaches would greatly assist in WB group comparisons.

To summarize, for clinical use where the aim is to keep additional effort to a minimum, the automatic scanner protocol has many advantages. However, the shortcomings of such an automated method need to be taken into account before choosing the type of input function used and should clearly be stated to provide readers with a clear picture of possible limitations. We herein validate the use of absolutely quantified values of glucose metabolism directly obtained from a state-of-the-art PET system with the arterial input function reference standard. When using the automatic vendor software, an inspection of VOI placement, IDIF and quantitative image estimates should be done before further use. Finally, when using protocols that require longer scan times or patient cohorts prone to involuntary movement, manual VOI definition, additional movement correction and manual modelling for results yield more reliable and robust results.