Background
Currently, the standard uptake value (SUV, in units of grams per milliliter), defined as the tracer concentration at a certain time point normalized to injected dose per unit body weight, is the only practical means typically used for quantitative evaluation of clinical [
18F-]fluorodeoxyglucose (FDG) positron emission tomography (PET) investigations. However, the SUV approach has several known shortcomings [
1‐
3] which affect the reliability of the SUV as a surrogate of the metabolic rate of glucose consumption. Among these are
1.
Susceptibility to errors in scanner calibration,
2.
Insufficient correlation between systemic distribution volume and body weight (leading to variants of the SUV approach using lean body mass (SUV
l
b
m
) [
4] or body surface area (SUV
b
s
a
) [
5] for normalization), and, consequentially, residual inter-study variability of the arterial input function (AIF) despite SUV normalization [
6].
3.
Time dependence of the SUV.
We have addressed the first two points in a recent publication [
7] and demonstrated that the standard tumor-to-blood uptake ratio (SUR) is superior to SUV as a surrogate parameter of
Km (the metabolic rate of FDG). The reason for this is that SUR can be shown to be linearly related to
Km if the AIF exhibits an essentially invariant shape across different investigations. Scale changes of the AIF (different blood SUVs) do not have any influence on SUR, whereas lesion SUV is directly affected by the latter.
While tumor SUR thus is not affected by inter-study variability of blood SUV, its use does not address another source of potentially serious variability, namely, insufficient standardization of the uptake time prior to scanning. Variability of the uptake period (i.e., variability of scan start time relative to the time of injection) is a persistent issue in clinical oncological PET [
8,
9]. This represents a well-known problem for meaningful SUV quantification since tumor SUV distinctly increases over time [
1,
2,
10,
11]. Especially affected are follow-up studies, where SUV changes of tumor lesions between consecutive studies are the relevant quantity (e.g., in the context of therapy response assessment), and scan time variability can lead to serious misinterpretation of the data.
We are aware of two studies addressing the question how to correct SUVs for variable scan time [
10,
11]. In [
10], a general SUV correction formula is proposed, but the emphasis of this paper lies on considerations regarding optimizing the imaging time, and the proposed semi-empirical correction formula is not fully evaluated. It contains, however, the important insight that the shape of the AIF enters the correction. A second investigation of the scan time variability effect and an empirical formula how to correct for it has been reported in [
11]. This investigation was restricted to breast cancer patients, and the proposed correction formula resulted from purely empirical observation of how SUV varied in these patients over time without explicitly considering the influence of the AIF shape. Here, the important observation was made that the rate of SUV change over time is approximately proportional to the magnitude of the SUV itself.
In the present work, we propose a generic method for correction of scan time variability effects on SUR (and SUV) which is based on observation and utilization of the specific shape invariance of FDG input functions and the consequences following from it when analyzing the Patlak equation. Especially, this provides the ability to map all measured values to a common reference scan time (e.g., 60 min p.i.) as long as FDG kinetics can be considered irreversible in the targeted tissue.
Starting from the Patlak equation, we first derive the correction formulas for SUR and SUV, respectively. The correction procedure is then evaluated in a group of nine patients with liver metastases of colorectal cancer, for which 15 dynamical investigations are available. These data enable assessment of the full dynamic AIF as well as comparison of the actually observed tumor SUV changes over time with the correction factors derived for two time frames selected from the dynamic data. Furthermore, we apply the correction procedure to an independent group of ten dual time point (DTP) whole body studies, where the full AIF is not known. Here, the correction is used to map SUR (and SUV) from the second to the first time point to investigate whether the corrected SURs and SUVs are concordant with those determined at the first time point as should be the case if the correction works well.
Discussion
Non-negligible variability of scan start relative to the time of injection is a persistent issue in clinical oncological PET [
8,
9]. In this work, we have developed and investigated a straightforward method to compensate for the spurious changes of SUR and SUV caused by this variability. We want to clarify from the outset that the proposed correction procedure is not intended to obviate the necessity for adequately standardized data acquisition as described, for example, in [
17]. Even though the presented scan time correction turns out to work very well even for substantial deviations from the targeted reference scan time, it is of course preferable to minimize the influence of the correction (with its inherent remaining uncertainties) by adhering to a standardized time p.i. as closely as practically possible.
The method was evaluated in two patient groups, first, in 22 liver lesions (15 scans) measured dynamically for 60 min and furthermore in 21 lesions (10 scans) assessed with DTP measurements. For each investigation, a reference time was selected from the available time frames (midpoint of the last time frame (T=55 min) for the dynamic studies and of the early DTP time frame (78.1 ± 15.9 min), respectively). SUR and SUV measured at the respective reference time were used as the ‘gold standard’ against which the corrected uptake values acquired at further time points were compared.
The main finding of this work is that correction for scan time variability (more precisely, correction for the accompanying spurious SUR and SUV changes) is feasible with surprisingly good accuracy and does only require an image-based determination of the arterial tracer concentration at the given scan time (rather than knowledge of the full AIF): group averaged differences between corrected and reference values typically only amount to about (3±7)% (or even less) and maximal deviations remain below ±15% in all cases. This is to be compared with group averaged deviations of up to about 50% (and maximal deviations of 65%) without correction.
Especially, the obtained accuracy can be considered sufficient to allow reliable comparison of lesion SURs and SUVs in follow-up studies during therapy response assessment, where typical inter-scan changes of 50% are considered as relevant (see e.g., [
18] and references therein) even if the scan time varies significantly between the successive scans. Moreover, compensation for scan time variability does offer the possibility to investigate whether smaller SUV changes (currently masked by scan time variability effects) could also provide relevant diagnostic information.
Two reasons can be identified for the observed residual differences
Δ SUR
0=SUR
0-SUR
ref between SUR measured at time
T (and corrected to chosen reference time) and the SUR actually measured at that reference time. The first reason, obviously, are limitations of the correction procedure and its underlying assumptions. The second one is the fact that we use SUR
ref as our
de facto ‘gold standard’ since the real ground truth SUR at that time point is not known. Therefore, the resulting
Δ SUR
0 is affected by statistical and systematic errors of both measured SUR values (SUR
0 and SUR
ref). In order to estimate the relative influence of both factors, we tentatively eliminated the first one by using Equation
4 instead of Equation
8 for correction and re-evaluation of the dynamic data using the individually correct
Vr resulting from Patlak analysis and the actual
Θ
T
resulting from integration of the individually measured dynamic AIF up to time
T. This procedure reduces the correction to application of the individually correct Patlak plot which thus would be ‘exact’ if the individual dynamic data points were perfectly following the Patlak equation. For
T=20 min, we find a reduction of the mean difference of SUR from (-2.9±6.6)% to (1.52±3.8)%, i.e., the mean value and the standard deviation of the difference are reduced by about 50%. The still remaining deviations can be attributed to the statistical and systematic errors of the two SUR measurements contributing to the correction. Therefore, it might be concluded that the accuracy of the presented scan time correction method actually is roughly twice as high than our results initially suggest: its inherent uncertainty is comparable to the statistical accuracy of the measured SUV and SUR values themselves.
The presented method rests on the empirical fact that the AIF after FDG bolus injection exhibits an invariant shape that, moreover, leads to an approximate proportionality between Patlak time
Θ and actual scan time
T (see Figure
1). This observation is mathematically equivalent to stating that starting very early after injection, the AIF can be described by a simple power law (see Figure
2). This specific, investigation-independent
Θ(
T) relation (Equation 7) ultimately leads to Equation
8. Thus, SUR correction back to the actually intended (reference) scan time only requires determination of the respective SUR, i.e., of tissue SUV as well as AIF level at the given scan time. It is then computable from the deviation of the given from the intended start time alone (using an estimate of the small parameter
whose precise value is not critical for the correction). An interesting observation in this context is the fact that the actual value of the time exponent
b in the power law used for describing the AIF does not explicitly enter the SUR correction (Equation 8) but only the SUV correction formula (Equation 10). The required value of
b was determined by a least squares fit of Equation
5 to the available group averaged dynamic AIF (see Figure
2) resulting in
b=0.313. Using this value of
b yielded a correction of SUV with an accuracy comparable to that of the SUR correction.
Strictly speaking, our dynamic liver scans only demonstrate that the performed parametrization of the AIF shape by an inverse power law and the stated value of the parameter
b holds up to about 60 min p.i.. Beyond that time range, no direct proof for this specific (and invariant) shape exists. However, the very good performance of the correction procedure not only for the dynamic liver metastases data measured up to
T=60 min but also for the DTP measurements (where the second scan occurred on average about 2 h p.i.) provides strong evidence that Equation
7 can be extrapolated to distinctly later times (with the same value for
b) without causing notable errors in the correction procedure. Moreover, the SUR correction (Equation 8) is not affected by the actual value of
b (and would be unaltered even in a constant infusion scenario corresponding to
b=0), while the SUV correction (Equation 10) is not critically sensitive to modest variations of
b. Good performance of the correction in the DTP group furthermore shows that the chosen value
(also derived from the dynamic AIF data of the liver metastases investigations) is applicable for all tumor entities included in the DTP investigations.
The chosen parametrization of the AIF by a power law allows a concise formulation of the correction formulas. While this parametrization is not usually applied for empirical modelling of AIF shapes (where sums of exponentials are more common), it is not only justified by our results but also consistent with the published data of Thie et al. which modeled
Θ(
T) (
Teq(
t) in their notation) as a third-order polynomial (Equation (
3) in [
10]). In the time range used by Thie et al. (up to 60 min p.i.), the non-linearity is very small and their relation quite well described by
Θ(
T)=1.67×
T which deviates by less than 15% from our result
m=1/(1-
b)=1.46. While the power law approach thus seems suggested by the data and is attractive due to its dependence on only two free parameters, we would like to emphasize that a different AIF modelling would also be possible and would yield comparable results regarding accuracy of the correction, presuming the parametrization is adequately fitting the measured AIF data. For example, Vriens and coworkers used three exponentials to model a population averaged AIF after bolus transition up to
T=45 min [
19]. Fitting the power law in Equation
5 to their parameterized AIF leads to
b≈0.34 which also is consistent with our result
b=0.313±0.030. Moreover, as already pointed out, adjusting the value of
b only affects (slightly) the SUV correction, but not the SUR correction.
Notable differences are to be expected, however, if different parameterizations of the AIF are extrapolated beyond the time range used in the fitting procedure. Especially, an extrapolated exponential model (derived from the limited early time range of dynamic measurements) will rapidly decrease to negligible concentration levels at later times which is not in accord with reality.
On the other hand, our extrapolation of the power law to the much later times used in the DTP measurements could be proven to yield very satisfactory results regarding SUR/SUV correction. Therefore, our assumption is that a power law with b≈0.3 remains valid for AIF modelling also at late times. The good performance of the correction procedure for the DTP data (acquired with a different scanner at much later times and also reconstructed differently) provides strong evidence for this conjecture. Despite these promising results, further investigations are desirable to further support the assumed power law behavior of the AIF at later times.
Ultimately, the proposed SUR (and SUV) correction rests on the ability to derive the lesion’s
Km from its measured SUR according to Equation
3. As already stated in [
7] this is not necessarily always correct, e.g., if inflammation is involved, this assumption no longer hold since the Patlak assumption of irreversible trapping is violated in this case. For that reason, we did not recommend a general conversion from SUR to
Km in [
7]. One example might be the finding in [
11] that a fraction of breast cancer lesions exhibited untypical time dependence of lesion SUV (very low and either essentially time independent or even slightly decreasing). While such behavior (including slightly decreasing SUV over time) still is compatible with irreversible kinetics if
Km is sufficiently small (due to the then dominant contribution of the reversible FDG pool to the PET signal which follows the decrease of the AIF), it also might be caused by actual deviations from irreversible kinetics. From a practical point of view, as far as lesions with a very low SUV are concerned (as was the case in [
11]), such deviations from irreversible transport have no real impact on the scan time correction for the simple reason that the absolute magnitude of the correction is small if the SUV itself is small, so, even if the correction would be erroneous in this situation, it would simply represent a small (erroneous) correction of a small SUV with no further practical consequences. We reiterate, however, that the correction procedure rests on the assumption of irreversible tracer kinetics and might lead to erroneous conclusions if this assumption is violated. On the other hand, the assumption has turned out to be valid for all tumor lesions we have investigated so far (notably the quite heterogeneous group of tumor entities present in our DTP data).
Altogether, scan time correction of SUR thus seems feasible whenever the tracer kinetics can be adequately described by a Patlak equation. Scan time correction of SUV is then possible as well, but somewhat more dependent on the applied AIF parametrization (explicit dependence on b). We believe it also would be worthwhile to investigate whether the procedure could be extended to other mostly irreversibly binding PET tracers (i.e., tracers for which influence of a possibly non-zero k4 rate constant in the standard reversible two-compartment model is negligible in the considered time range) if SUV-based approaches are applied since they would be affected by scan time variability in a comparable way.
Finally, it is important to point out the fact that the time dependency of SUR is distinctly larger than that of SUV which makes scan time correction even more important for the former. The increased time dependence is easily explained by observing that SUR is defined as the ratio of (increasing) SUV and (decreasing) arterial blood concentration. For our data, the difference between SUR at the actual scan time
T and SUR at the reference time was up to
δ SUR
T
≈60
%. The SUV difference was lower, up to
δ SUV
T
≈40
%, but still far from negligible. Moreover, the scan time correction of SUV does of course not reduce the independent substantial influence of inter-study variability in blood concentration on the resulting SUV [
7]. This distinct disadvantage of SUV relative to SUR persists, and we believe that a transition from SUV-based (usually corrected neither for scan time nor for arterial blood concentration variability) to SUR-based evaluation (including correction for scan time variability as proposed in the present work) could offer distinct advantages for quantitative oncological FDG PET.
Authors’ contributions
JVDH had the initial idea for time correction of SUR and SUV, performed part of the data analysis, and wrote part of the manuscript. AL and GS performed part of the data analysis. JM, LO, JP, and JK provided intellectual input and reviewed the manuscript. BBB selected the patient studies and performed the lesion delineation. FH performed part of the data analysis and wrote part of the manuscript. All authors read and approved the final manuscript.