Background
In clinical oncology, positron-emission tomography (PET) is a valuable tool allowing guidance of treatment on a per-patient basis [
1]. Clinical decision-making using PET-CT is commonly limited to visual analysis, where local disease and the presence of nodal or distant metastases is evaluated [
2,
3]. However, since PET is an inherently quantitative technique, it may also be used for quantitative assessment of tumor metabolic, proliferative, or drug targeting characteristics [
1,
4,
5].
For quantitative PET-CT to be of practical clinical utility, metrics need to be easily extracted from static whole-body PET-CT images as performed in routine clinical practice. To this end, standardized uptake values (SUV) are typically used as simplified semi-quantitative measures of tracer uptake [
6]. However, pharmacokinetic modeling using dynamic PET-CT acquisitions with arterial or venous blood sampling is an essential first step to technically validate the clinical use of these simplified metrics as biomarkers of, e.g., response to treatment [
4,
5,
7,
8].
As is well known, quantification of tracer distribution on PET-CT scans is hampered by several sources of error. Among these are attenuation, Compton scatter, random coincidences, and decay, all accounted for by contemporary image reconstruction algorithms. However, due to the inherently limited spatial resolution of PET-CT, acquired images still suffer from partial-volume effects [
9]. Partial-volume effects lead to spill-in and spill-out of measured activity distributions, generally resulting in net underestimations of tracer uptake, the extent of which depend on tumor size, shape, and contrast [
9]. Hence, partial-volume correction (PVC) is needed for accurate quantification, especially for small and/or heterogeneous lesions [
9‐
12].
In oncological studies, PVC has been predominantly applied to static PET-CT images (in contrast with brain [
13‐
22] or cardiac [
23,
24] PET imaging). However, in dynamic acquisitions, the activity spill-over in and from tumors due to partial-volume effects may vary over time. The impact of PVC on tumor kinetic parameter estimates could therefore differ from its impact on simplified measures of uptake. Consequently, it may not only affect absolute quantitative reads but also validation of simplified parameters for clinical implementation.
The present study aims to evaluate the impact of frame-wise parametric PVC in dynamic PET-CT studies on tumor kinetic micro- and macroparameter estimations, and evaluate the correlation between its effect on kinetic parameters and simplified metrics. Secondly, PVC’s effect on technical validation of simplified 18F-fluorothymidine (18F-FLT) PET-CT metrics as biomarkers of response to treatment of non-small cell lung cancer (NSCLC) with tyrosine kinase inhibitors (TKI) will be investigated.
Discussion
In the present study, we evaluated the impact of frame-wise parametric PVC on tumor kinetic parameter estimation derived from dynamic PET-CT scans and the resulting effect on validation of simplified metrics. PVC significantly increased both tumor micro- and macrokinetic parameters, and we observed that partial-volume effects varied over time due to blood pool activity and changing tumor contrast. Hence, the effect of PVC on kinetic parameter estimates was not in full concordance with its effect on simplified metrics (SUV and TBR), and as a consequence, PVC was found to affect the validation of SUV using VT both for single measurements and as biomarker of treatment response to a small extent (albeit non-significantly).
Application of PVC in oncologic dynamic PET-CT studies is scarce. Mankoff et al. (2003) applied PVC in dynamic FDG-PET of breast cancer patients using a simple method with recovery coefficients, assuming lesions are spherical with homogenous tracer distributions [
29]. They observed that applying PVC in response measurements reduced changes in metabolic rate of FDG and blood flow of responding patients, reducing significance of parameter changes (albeit still statistically significant). By using this method, however, kinetic parameters were solely corrected for (changes in) tumor size, and no correction for spill-in from blood pool structures and/or heterogeneous tumor background was applied. In 2007, Teo et al. validated the use of iterative deconvolution as an image-based PVC method not requiring anatomical segmentation or knowledge of lesion size, and suggested its potential application in kinetic modeling, which to the best of our knowledge has not been performed to date for oncologic PET-CT [
30].
Both tumor macroparameters
VT and BP
ND, and microparameter
K1 significantly changed after application of PVC. This corresponds with results from applications of PVC in brain dynamic PET studies, where similar increases in kinetic parameter estimations have been observed when applying PVC in the case of activity spill-out [
19‐
21,
31]. Interestingly, the effect of PVC on kinetic parameter estimates was poorly (albeit significantly) correlated with its effect on simplified measures. As previously described [
9], the effect of PVC on SUV of (hotspot) lesions on static PET-CT scans is straightforward: an expected net increase in activity, mainly dependent on lesion size (and, in lesser extent, shape and local contrast). This can be seen in Fig.
2, where change in SUV after PVC is highly (inversely) correlated to tumor volume, while the kinetic parameter estimations are not. This illustrates that impact of PVC on tumor kinetic parameter estimation is more complex, as seen in Fig.
1 which displays the non-linear temporality of partial-volume effects for a typical mediastinal lymph node metastasis. Here, an early spill-in of activity due to blood pool proximity is noted, with increasing activity spill-out afterwards as tumor uptake increases and background activity decreases. Hence, across lesions, the effect of PVC on kinetic parameters may differ depending not only on size, but as well on the presence of proximate high activity structures, rate of tracer uptake during the scan, and background activity.
For quantification of functional tumor characteristic on PET-CT in clinical practice, a simplified quantitative method is necessary, obviating the need for complex and extended dynamic image acquisitions, need for blood sampling, and facilitating the possibility of whole-body acquisitions. To this end, per radiotracer and cancer type simplified metrics needs to be technically validated by pharmacokinetic modeling using dynamic PET-CT [
4]. In the current study, the effect of PVC on kinetic parameter estimates was different from its effect on simplified metrics, which explains why it might affect validation of these simplified metrics (using
VT). We observed a trend that PVC increased correspondence of SUV with
VT in single measurements (correlations improving from 0.82 to 0.90) and as a biomarker of treatment response (correlations improving from 0.90 to 0.95 at 7 days and from 0.79 to 0.88 at 28 days after treatment start). However, confidence intervals of these correlations overlapped, which might at least partly be due to the sample size (inherent to this type of study), and therefore these differences are not statistically significant. Therefore, while PVC is mandated to acquire accurate quantitative reads, it only increases correspondence of kinetic parameters with simplified metrics to a small extent on a cohort level. This indicates that the impact of image resolution on technical validation of simplified metrics of
18F-FLT as biomarkers of response to TKI might be small, and that PET images without PVC seem non-inferior for this purpose. It should be noted that for response assessment to treatments that affect tracer kinetics and blood pool activity to a larger extent than TKIs and for other cancer types more affected by spill-in (e.g., prostate cancer lesions with urinary tract proximity), PVC may have a larger impact on validation of simplified metrics.
Spill-out due to PVE will result in overestimation of metabolic tumor volumes, which increases the underestimation of true tracer uptake since background activity is included [
11]. A parametric PVC method may therefore theoretically reduce inaccuracies in delineation. However, iterative deconvolution has been proposed with use of VOIs defined on uncorrected images, due to the expected propagation of image noise after PVC [
30]. We evaluated the impact of delineation on deconvoluted images with HYPR denoising, and found not only substantial decreases in MATVs (Fig.
3) but also an increase in PVCs effect on kinetic parameter estimates (Additional file
1: Table S3). Nonetheless, our previous study demonstrated that the reduction in MATV after PVC may not necessarily lead to more accurate definition of tumor volumes [
11].
In brain PET studies, frequently a small vessel such as the carotid artery needs to be utilized for IDIF generation. This mandates PVC due to the small artery diameter [
32,
33]. In this study on thoracic oncological PET-CTs, the ascending aorta, a large vessel, was used for IDIF generation. We noted that PVC introduced negligible differences in IDIF area under the curves, and that without denoising this introduced small but significant differences in kinetic parameter estimates (Additional file
1: Table S2). However, since HYPR denoising using a single composite image (providing maximum noise reduction) appeared to completely mitigate this effect, the effect of PVC on these input functions seems to be based on PVC-induced noise-propagation. Therefore, when input functions derived from large blood pool structures are used, PVC is preferably avoided to evade noise-induced inaccuracies in kinetic parameter estimates (assuming no spillover from nearby high activity structures).
Iterative deconvolution algorithms are known to propagate image noise, which may necessitate denoising methods to be applied to preserve image quality. Several approaches have been proposed, such as wavelet-based denoising for static PET-CT and HYPR denoising for dynamic acquisitions, respectively [
26,
34]. We observed that HYPR needs to be optimized for tracer kinetics using a moving composite image, since when applied using a single composite image (maximal denoising) it seems to lose the temporal dynamic course of the PVC (Fig.
1). Including HYPR
moving resulted in very similar outcomes compared to PVC alone, and slightly mitigated the increase in kinetic parameter estimates after PVC. The latter may not only be attributed to reduced statistical noise but also to some smoothing effects inherent to the algorithm. Also, at late time frames, it had no effect on intratumoral COV% (Additional file
1: Figure S1). This might be explained by the high tumor contrast and high count number (due to the long frame duration), as Golla et al. previously demonstrated [
21]. The increase in COV% at late time frames thus seems to be a resultant of increased intratumoral heterogeneity by PVC itself. Therefore, in region-based non-linear regression analyses, the impact of PVC-induced increased image noise on kinetic parameter estimation seems negligible. However, it may have significant impact when tumors are analyzed on a parametric level.
While the presence of PVE and the consequent need for PVC are well recognized, to date PVC has rarely been applied in oncological PET studies. This may be because to date there is no consensus on the optimal correction strategy and data yielded from application of PVC does not seem to have triggered routine clinical application [
12,
35]. Our study now demonstrates that PVC should not only be performed in future regular static PET-CT studies, but in dynamic PET-CT studies as well, also when simplified quantitative metrics are validated for clinical applications. If not applied, small lesions should preferably be excluded from analyses, as recommended and performed in previous studies using a 2–3-cm-diameter cut-off to avoid PVE [
36,
37]. Still, our data demonstrate that lesions above these size thresholds are also affected by PVE (Fig.
2).
Only data from
18F-FLT PET-CT was used. However, the current dataset from a widely used whole body TOF PET-CT scanner allowed for both kinetic modeling and extraction of simplified parameters per lesion, at time points used in clinical practice due to the long acquisition time (0–60 min post-injection). Also, the dataset included both large and small lesions, both nearby and remote from large blood pool structures. Additionally, it facilitated evaluation of PVCs effect on validation of simplified parameters both in single measurements and during systemic treatment. Since we have demonstrated the significant effect of PVC in kinetic parameter estimation, future dynamic PET studies focusing on other PET-tracers in small tumors (e.g., PSMA-ligand tracers in prostate cancer metastases) should apply PVC as a similar (or larger) impact of PVC may be expected. In the current study, no correction was made for potential motion blurring effects, which is another factor possibly affecting accuracy of kinetic parameter estimations [
38]. Efforts should be made to incorporate both PVC and motion correction methodologies simultaneously for dynamic PET studies. Also, the impact of PVC on parametric kinetic analyses of oncologic dynamic PET warrants further investigation, which will require HYPR denoising to be optimized for this purpose.
Acknowledgements
We thank the participants for their participation in this study. We acknowledge the staff of the department of Radiology and Nuclear Medicine of the Amsterdam UCM (location VUmc) for their work in tracer production and acquisition of the scans made for this study. We acknowledge the members of the QuIC-ConCePT Consortium, whose participants include AstraZeneca, the European Organization for Research and Treatment of Cancer (EORTC), Cancer Research U.K., the University of Manchester, Westfälische Wilhelms-Universität Münster, Radboud University Nijmegen Medical Center, Institut National de la Santé et de la Recherche Médical, Stichting Maastricht Radiation Oncology “Maastro Clinic,” VUmc Amsterdam, King’s College London, Universitair Ziekenhuis Antwerpen, Institute of Cancer Research–Royal Cancer Hospital, Erasmus Universitair Medisch Centrum Rotterdam, Imperial College of Science Technology and Medicine, Keosys S.A.S., Eidgenössische Technische Hochschule Zürich, Amgen NV, Eli Lilly and Company Ltd., GlaxoSmithKline Research & Development Limited, Merck KGa, Pfizer Limited, F. Hoffmann–La Roche Ltd., and Sanofi-Aventis Research and Development.