Background

Positron emission tomography using [18F]fluorodeoxyglucose (18F-FDG PET) imaging in oncology patients allows physicians to quantify the increased glycolysis of cancer cells [1]. In clinical routine, a tracer uptake index is easily available and thus widely used, namely, the standardized uptake value (SUV) [2, 3]. However, many factors can influence the SUV outcome such as the uptake time, as reported for example in lung tumors [4]. This is the reason why, besides the SUV index, different quantitative parameters that may be obtained from kinetic model analyses (KMAs) have been implemented in a number of studies investigating various tissues [5,6,7,8,9,10,11,12,13]. These kinetic parameters more accurately describe the tracer trapping and may be useful to better characterize different tumor types or assess treatment response [14]. The KMAs both require a dynamic acquisition over the tissue of interest to obtain its time-activity-curve (TAC) and a serial blood sampling to estimate the so-called input function (IF, i.e., 18F-FDG blood TAC). Among these KMAs, Patlak’s analysis is usually considered as a gold standard that provides the 18F-FDG net influx constant (i.e., the uptake rate constant, Ki) from a linear fitting of graphical data [7]. However, it assumes an irreversible tracer trapping, a well-identified drawback since numerous studies have shown trapping reversibility in various tissues, either under physiological or pathological conditions [6, 9,10,11, 15].

Assuming that there may be a slow loss of the trapped tracer to the blood, i.e., a reversible trapping, Patlak and Blasberg derived a generalized non-linear equation including a release rate constant (kb) in an exponential term [8]. This non-linear equation may be addressed by using an analytical approach, leading to a three-compartment three-parameter KMA (3P-KMA). 3P-KMA has been applied to healthy human lung and liver, allowing assessment of both Ki (in mL min−1 mL−1), kb (in min−1), and fraction of free 18F-FDG in blood and interstitial volume (Vb; no unit; < 1; also called total blood volume distribution) [12, 15]. It relies on an analytical solution  of the non-linear Patlak’s equation that requires to use an IF as a sum of exponentially decreasing functions and up to three functions may usually describe the 18F-FDG IF [16, 17]. Then, it leads to a non-linear formula that is used to fit the experimental tissue TAC by simply adjusting the three above-mentioned kinetic parameters, i.e., without any tissue TAC data transform.

To the best of our knowledge, comparison between non-linear fitting by 3P-KMA and linear fitting by standard Patlak’s analysis that assumes an irreversible tracer trapping has not been reported so far, whatever the tissue either under physiological or pathological conditions. Therefore, the primary aim of this study was to make this comparison in a series of lung cancer patients that was previously acquired [13]. Additionally, Ki, kb, and Vb outcomes obtained from 3P-KMA were compared to those obtained from a three-compartment five-parameter KMA (5P-KMA) that is usually considered as a reference model when tracer trapping is reversible. Actually, 5P-KMA provides four kinetic (micro)parameters (and Vb) from which Ki and kb may be computed, whereas the non-linear fitting of the 3P-KMA provides Ki and kb without any additional computing.

Methods

Patients

Dynamic data of 21 patients (8 females, 13 males, 71 years old on average, range 40–86) with non-small cell lung cancer obtained from a previous prospective study were retrospectively analyzed [13]. All patients who were enrolled in the prospective study provided written informed consent before participating in it, and the further retrospective study received the approval of the ethics committees of our teaching hospitals. The patients’ mean weight and height were 68 kg (range, 50–85) and 169 cm (range, 150–180), respectively. After 6 h of fasting before the tracer injection, the preinjection average plasma glucose concentration was 1.08 g L−1 (range, 0.87–1.28). The lesion mean size was 31.8 mm (range, 14.7–52.2).

PET imaging and data processing

PET imaging procedure has been previously described in details [13]. Briefly, a low dose CT scan was performed (75 mA, 120 kV, pitch 0.938, rotation time 0.5 s) for attenuation correction of PET emission data and for morphologic information. Then, after an intravenous bolus injection of 18F-FDG (mean 237 MBq; range, 134–507) in a cannula previously inserted in the vein of the arm, a 3D thorax dynamic list-mode acquisition protocol was started lasting 60 min (Gemini GXL, Philips Medical System, Cleveland, USA; no respiratory gating). Images were reconstructed using the iterative method RAMLA LOR-3D, with a 144 × 144 matrix and pixel size of 4 × 4 × 4 mm3. In particular, this dynamic acquisition provided 11 frames of 5 min each, leading to 11 data points of the experimental 18F-FDG IF and of the experimental cancer tissue TAC, ranging 7.5–57.5 min post-injection. For determining the experimental 18F-FDG IF, in each patient, a volume of interest (VOI) was drawn over the descending thoracic aorta in each frame of the dynamic acquisition yielding an intermediate 18F-FDG blood TAC (i.e., intermediate IF). Then, the final IF was obtained through a calibration of the intermediate IF with the 18F-FDG plasma value measured in a venous blood sampling performed at 45 min post-injection, that is, when an equilibrium is reached between 18F-FDG concentration in arterial and vein blood [3, 16]. VOIs for IF and tissue TAC were semi-automatically placed over three consecutive slices to include the five hottest voxels within the VOI.

Implementing Patlak’s analysis, 3P-KMA, and 5P-KMA

Assuming that there may be a slow loss of the trapped tracer to the blood and when the analysis remains limited to data collected for the period t > t* after injection, that is, when the reversible compartments are in effective steady state with the blood plasma, Patlak and Blasberg derived a non-linear equation including a release rate constant (kb) [8]:

$$ {\mathrm{A}}_{\mathrm{T}}\left(\mathrm{t}\right)/{\mathrm{A}}_{\mathrm{p}}\left(\mathrm{t}\right)=\left[\mathrm{Ki}{\int}_{\mathrm{o}}^{\mathrm{t}}{\mathrm{A}}_{\mathrm{p}}\left(\uptau \right){\mathrm{e}}^{\hbox{-} {\mathrm{K}}_{\mathrm{b}}\left(\mathrm{t}\hbox{-} \uptau \right)}\mathrm{d}\uptau \right]/\left[{\mathrm{A}}_{\mathrm{p}}\left(\mathrm{t}\right)\right]+{\mathrm{V}}_{\mathrm{b}} $$
(1)

AT(t) (in kBq mL−1) is defined as the total tracer activity at time t per tissue volume unit that includes both trapped tracer and free tracer in the blood and interstitial volumes. Ap(t) (in kBq mL−1) is the blood activity at time t per blood volume unit, that is, the 18F-FDG IF.

In each patient, Patlak’s graphical analysis was implemented from Eq. 1, setting kb = 0. Eleven cancer tissue TAC data points and the corresponding 11 data points of the experimental 18F-FDG IF, ranging 7.5–57.5 min post-injection, were used. The lower limit of 7.5 min for this range was chosen in order to limit the analysis to data collected for the period t > t* after injection, as required by Eq. 1 validity [7, 8]. Ki was determined as the slope of the linear fitting of the Patlak’s plot showing AT(t)/Ap(t) versus the ratio of time integral of the right hand side of Eq. 1 to Ap(t), i.e., the so-called stretched time.

In each patient, 3P-KMA was implemented by first fitting the 11 data points of the experimental 18F-FDG IF with a three exponentially decreasing function derived from Hunter’s results, after data were uncorrected for the 18F physical decay. Hunter’s results were used, and not Vriens’ ones as in previous studies, because the former were established with blood sampling performed at 55 min post-injection, in comparison with 25 min for the latter [12, 15,16,17]:

$$ {\mathrm{A}}_{\mathrm{p}}\left(\mathrm{t}\right)={\mathrm{A}}_0\times \left[8.20\times \exp \left(\hbox{-} 9.3363\times \mathrm{t}\right)+1.17\times \exp \left(\hbox{-} {\upalpha}_2\times \mathrm{t}\right)+\exp \left(\hbox{-} {\upalpha}_3\times \mathrm{t}\right)\right] $$
(2)

In Eq. 2, the amplitude ratios of the three exponential functions and the time constant of the first exponential function (uncorrected for physical decay) were available from Hunter’s results [16]. In each patient, A0 (leading to virtual initial IF amplitude), α2, and α3 (time constants of the second and third exponential functions, uncorrected for physical decay) were obtained by fitting the experimental 18F-FDG IF data points (XLSTAT Microsoft; Levenberg-Marquardt algorithm). Then, in each patient, a formula was established by analytically solving integral of Eq. 1 and by using the fitted three-exponential IF of Eq. 2 involving kb [12, 15]:

$$ {\displaystyle \begin{array}{ll}{\mathrm{A}}_{\mathrm{T}}\left(\mathrm{t}\right)=\mathrm{Ki}\times {\mathrm{A}}_0& \times \Big\{8.20\times \left[\exp \left(\hbox{-} 9.3363\times \mathrm{t}\right)\hbox{-} \exp \left(\hbox{-} \left(\uplambda +{\mathrm{k}}_{\mathrm{b}}\right)\times \mathrm{t}\right)\right]/\left[\left(\uplambda +{\mathrm{k}}_{\mathrm{b}}\right)\hbox{-} 9.3363\right]\\ {}& +1.17\times \left[\exp \left(\hbox{-} {\mathrm{a}}_2\times \mathrm{t}\right)\hbox{-} \exp \left(\hbox{-} \left(\uplambda +{\mathrm{k}}_{\mathrm{b}}\right)\times \mathrm{t}\right)\right]/\left[\left(\uplambda +{\mathrm{k}}_{\mathrm{b}}\right)\hbox{-} {\mathrm{a}}_2\right]\\ {}& +\left[\exp \left(\hbox{-} {\mathrm{a}}_3\times \mathrm{t}\right)\hbox{-} \exp \left(\hbox{-} \left(\uplambda +{\mathrm{k}}_{\mathrm{b}}\right)\times \mathrm{t}\right)\right]/\left[\left(\uplambda +{\mathrm{k}}_{\mathrm{b}}\right)\hbox{-} {\mathrm{a}}_3\right]\Big\}\\ {}& +{\mathrm{V}}_{\mathrm{b}}\times {\mathrm{A}}_{\mathrm{p}}\left(\mathrm{t}\right)\end{array}} $$
(3)

In Eq. 3, the 18F physical decay constant is λ, and Ki, kb, and F were obtained in each patient by fitting the 18F-FDG tissue TAC (XLSTAT, Microsoft; Levenberg-Marquardt algorithm), ranging 7.5–57.5 min post-injection, uncorrected for 18F physical decay. Note that previous studies used 18F-FDG tissue data obtained at late dynamic PET imaging, i.e., beyond 2 h after injection, in comparison with the current ones obtained at early imaging (7.5–57.5 min post-injection) [12, 15]. However, the rationale for deriving Eq. 3 remains identical, whatever the time of acquisition (>t* after injection).

In each patient, the 5P-KMA model was implemented on PMOD software by using the whole experimental IF and tissue TAC data points acquired from injection, with very short frames including the bolus injection (version 3.0; PMOD Technologies, Switzerland) [13]. The 5P-KMA model can provide four kinetic rate constants, i.e., K1, k2-3-4 and Vb: K1 and k2 account for forward and reversed transport between blood and reversible compartment, and k3 and k4 account for forward and reversed transport between reversible and trapped compartment, respectively [12, 15]. The rate constants Ki and kb may be computed from K1, k2-3-4, as:

$$ \mathrm{Ki}={\mathrm{K}}_1\ {\mathrm{k}}_3/\left({\mathrm{k}}_2+{\mathrm{k}}_3\right) $$
(4)
$$ {\mathrm{k}}_{\mathrm{b}}={\mathrm{k}}_2\ {\mathrm{k}}_4/\left({\mathrm{k}}_2+{\mathrm{k}}_3\right) $$
(5)

Statistical analysis

A normal distribution of the α2 and α3 values (Eq. 2) in the current study and in Hunter’s study could not be clearly showed for each IF time constants; therefore, comparisons between the two studies were made by means of non-parametric Mann-Whitney’s test (GraphPad Prism 6 software; two-tailed; 95% confidence level). Bland-Altman analysis was used for Ki comparison between 3P-KMA and Patlak’s analysis, as well as for further comparisons between 3P-KMA and 5P-KMA, and between 5P-KMA and Patlak’s analysis (GraphPad Prism 6 software; 95% confidence level) [18].

Results

Figure 1 shows an IF fitting in a typical patient. Values of A0, α2, and α3 are presented for each patient in Table 1. Range of IF fitting correlation coefficients over the patient series was 0.989–0.999 (mean 0.996). The α2 and α3 values found in the current study were significantly lower than those of Hunter’s study (after removing the decay correction): P < 0.0001 for the two comparisons.

Fig. 1
figure 1

Typical IF fitting (patient 9 in Table 1; R = 0.989; P < 0.001). PET data (square) are uncorrected for 18F physical decay

Table 1 Fitting results in each patient for Patlak’s analysis, 3P-KMA and 5P-KMA (A0 in kBq mL−1; α2, α3, k b k2-4 in min−1; Ki and K1 in mL min−1 mL−1; Vb no unit)

Figures 2 and 3 show linear Patlak’s fitting and non-linear 3P-KMA in the same patient as in Fig. 1. Values of Ki obtained from Patlak’s analysis versus Ki, kb, and Vb obtained from 3P-KMA are presented for each patient in Table 1. The range of correlation coefficients for Patlak’s and 3P-KMA fitting over the patient series was 0.971–0.999 and 0.766–0.998 (mean 0.990 and 0.958), respectively. A significant correlation was found between correlation coefficients of IF fittings and those of 3P-KMA fittings (R = 0.631; P < 0.01; graph not shown).

Fig. 2
figure 2

Patlak’s analysis performed in patient 9 (Table 1): y = 0.0403x + 0.0059 (R = 0.999; P < 0.001), indicating that the Ki of patient 9 is 0.0403 mL.min-1.mL-1

Fig. 3
figure 3

Typical 3P-KMA fitting of 18F-FDG tissue TAC (patient 9 in Table 1; R = 0.998; P < 0.001). PET data (square) are uncorrected for 18F physical decay

3P-KMA Ki was found to be strongly correlated with Patlak’s Ki (R = 0.995; P < 0.001; graph not shown). Figure 4 shows the comparison between 3P-KMA Ki and Patlak’s Ki in the manner of Bland-Altman [18]. 3P-KMA Ki was significantly greater than Patlak’s analysis Ki: Ki ratio (i.e., 3P-KMA/Patlak) which was 1.060 ± 0.040 on average (95% confidence limits), with 95% limits of agreement of 0.171. When patients with kb > 0 were excluded (n = 8; Table 1), 3P-KMA Ki was no more significantly greater than Patlak’s Ki: Ki ratio which was 1.014 ± 0.030 on average (95% confidence limits), with 95% limits of agreement of 0.098. A strong correlation was found between Ki ratio and kb (Fig. 5; R = 0.801; P < 0.001).

Fig. 4
figure 4

Ki ratio of 3P-KMA/Patlak against mean. Ki ratio was 1.060 ± 0.040 on average (central dashed line; 95% confidence limits not shown), with 95% limits of agreement of 0.171 (upper and lower dashed lines)

Fig. 5
figure 5

Ki ratio of 3P-KMA/Patlak versus kb: y = 43.679x + 1.022, R = 0.801; P < 0.001

Values of Ki, kb, and Vb obtained from 3P-KMA and values of K1, k2-3-4, Vb, Ki, and kb (computed from Eqs. 4 and 5) obtained from 5P-KMA are presented in Table 1. Range of correlation coefficients for 5P-KMA fitting over the patient series was 0.977–0.999 (mean: 0.989). 3P-KMA Ki was found to be strongly correlated with 5P-KMA Ki (R = 0.989; P < 0.001). No significant difference was found between 3P-KMA Ki and 5P-KMA Ki: Ki ratio (i.e., 3P-KMA/5P-KMA) which was 1.017 ± 0.054 on average (95% confidence limits), with 95% limits of agreement of 0.230. No significant difference was found between 5P-KMA Ki and Patlak’s Ki: Ki ratio (Table 1) (i.e., 5P-KMA/Patlak) which was 1.056 ± 0.074 on average (95% confidence limits), with 95% limits of agreement of 0.317. 3P-KMA kb was found to be significantly correlated with 5P-KMA kb (R = 0.60; P < 0.01). No significant difference was found between 3P-KMA kb and 5P-KMA kb: kb difference (i.e., 3P-KMA minus5P-KMA; kb ratio is not allowed since division by zero is not allowed) which was 0.00041 ± 0.00083 min−1 on average (95% confidence limits), with 95% limits of agreement of 0.00359 min−1. No significant correlation was found between 3P-KMA Vb and 5P-KMA Vb (R = 0.12).

Figure 6 shows the tissue TAC and the part of trapped tracer and of free tracer in blood and interstitial volume (Eq. 3), by using mean values for IF and for 3P-KMA parameters (Ki, kb, Vb) obtained over the current lung cancer series.

Fig. 6
figure 6

(Full) Average tissue TAC obtained from mean values of the kinetic parameters (Ki, kb, Vb) found over current lung cancer series; (dotted) trapped tracer TAC; (dashed) TAC of free tracer in blood and interstitial volume. Peak time for tissue TAC and trapped tracer TAC is 84 and 88 min, respectively. Data are uncorrected for 18F physical decay

Discussion

In each patient, 11 data points of the 18F-FDG IF were fitted by using a three-exponential decreasing function. These data points ranged 7.5–57.5 min post-injection, that is, after an equilibrium has been reached between compartments in order to satisfy the Patlak’s condition t > t*. This function was derived from Hunter’s results, of which A0, α2, and α3 were obtained by fitting (Eq. 2) [16]. Indeed, the relative part of each exponential function to the IF area-under-curve (i.e., the total number of molecules that are available to the tissues after injection) is 1.14, 9.62, and 89.24% (by using the mean value of α1 by Hunter and of α2 and α3 reported in Table 1), respectively. In other words, the part of the first exponential function in the whole IF, which mainly covers the IF peak, is very limited, suggesting that the mean value of α1 reported by Hunter may be used in each individual [16]. Comparison between the fitted IFs of the current study and those reported by Hunter et al. shows that the former α2 and α3 values were significantly lower than the latter ones (P < 0.0001) [16]. The IF fitting correlation coefficients were high (range, 0.989–0.999; P < 0.001; Fig. 1). The major role of a reliable analytical IF as a sum of exponential functions for implementing 3P-KMA (Eq. 3) is emphasized by the significant correlation between IF fitting correlation coefficients and those of 3P-KMA (R = 0.631; P < 0.01). It is noteworthy that, although the current study with 18F-FDG used a three-exponential decreasing function, 3P-KMA may also be efficient with either a mono- or a bi-exponential IF depending on the tracer. Moreover, the ways the exponentially decaying IF can be obtained may be various: either from arterial or venous blood sampling, or image-derived, or from population-based IF models possibly scaled to later dynamic measurements on blood pool ROIs [3, 19].

3P-KMA Ki was found to be 6.0% greater than Patlak’s Ki, on average, with reasonable 95% limits of agreement of 17.1%, according to Bland-Altman analysis (Fig. 4). Moreover, when patients with kb > 0 were excluded, 3P-KMA Ki was no more significantly greater than Patlak’s Ki, with 95% limits of agreement of 9.8%. These findings were in agreement with Patlak and Blasberg’s comment suggesting that, in case of reversible trapping (kb > 0), linear fitting of a concave curve rather than that of a (true) linear one (kb = 0), results in an underestimation of the slope and hence to Ki underestimation [8]. These findings also suggest that 3P-KMA and Patlak’s analysis may be used interchangeably to assess 18F-FDG uptake in lung cancer lesions. However, the strong correlation between Ki ratio (i.e., 3P-KMA/Patlak) and kb (and hence, between Ki underestimation by Patlak’s analysis and kb; Fig. 5) suggests that 3P-KMA may be more appropriate than Patlak’s analysis to accurately assess Ki in various 18F-FDG-positive cancer lesions, possibly showing greater trapping reversibility than lung cancer lesions [9,10,11].

No significant difference over the current series was found between 3P-KMA Ki and 5P-KMA Ki and between 3P-KMA kb and 5P-KMA kb. A further comparison strengthens the current findings, between the current 3P-KMA outcomes and published ones by Dimitrakopoulou-Strauss et al., who implemented 5P-KMA in nine patients with lung tumors: Ki = 0.0414 ± 0.0288 min−1 (SD) and kb = 0.0009 ± 0.0016 min−1 (SD) for 3P-KMA (Table 1) versus the mean value of 0.0304 and 0.0009 min−1 for 5P-KMA by Dimitrakopoulou-Strauss et al., respectively [11]. However, it should be emphasized that the measurement uncertainty of the 3P-KMA outcomes may be expected to be lower than that of the 5P-KMA one because Ki and kb from 5P-KMA has to be computed by using three independent kinetic (micro)parameters (Eqs. 4 and 5). As a result, the measurement uncertainty of Ki and kb from 5P-KMA combines that of the three (micro)parameters, whereas the measurement uncertainty of Ki and kb from 3P-KMA may be obtained without any further combination [20]. (Note that MU of 3P-KMA outcomes was not available from XLSTAT that did not allow a possible comparison with MU of 5P-KMA outcomes.) The proposed line of argument may be associated with Galli et al.’s results showing that close values of Ki may be computed from different set of (micro)parameter values [3]. It may also explain why, unlike for the comparison between 3P-KMA Ki and Patlak’s Ki, no significant difference was found between 5P-KMA Ki and Patlak’s Ki and, hence, that the Ki underestimation by Patlak’s analysis was not revealed by the latter comparison. It may also be illustrated by the comparison of limits of agreement of 17.1 versus 31.7% that were found for the comparison between 3P-KMA and Patlak’s analysis versus the comparison between 5P-KMA and Patlak’s analysis, respectively. Furthermore, no significant correlation was found between 3P-KMA Vb and 5P-KMA Vb (R = 0.12), and the Vb mean value over the current series was found to be 0.16 ± 0.09 (SD) and 0.05 ± 0.03 (SD), respectively (Table 1). Consistently with the above-proposed comparison for Ki and kb, comparison of the current 3P-KMA Vb value of 0.16 ± 0.09 (SD) with that of 0.17 ± 0.07 (SD) previously published for 5P-KMA by Dimitrakopoulou-Strauss et al. further strengthens the findings of the current study [11].

Unlike Patlak’s analysis, the 3P-KMA approach allows expressing the whole tissue TAC as an analytical formula (Eq. 3). Therefore, at each time point, it is possible to compare it to that of its two components, that is, to the trapped tracer TAC and to the free tracer TAC. This comparison is shown in Fig. 6 by using mean values for IF and for 3P-KMA kinetic parameters that were obtained over the current lung cancer series (Table 1). Furthermore, this graph shows that the (mean) peak time for tissue TAC and trapped tracer TAC is 84 and 88 min, which could serve as landmarks to determine the optimal injection-acquisition time delay in clinical practice.

A limitation of the study is that, although the current 3P-KMA results obtained for Ki, kb, and Vb were in agreement with previous literature results by Dimitrakopoulou-Strauss et al., the kb/Ki ratio was low, about 2% on average in the current lung tumor series (Table 1) [11]. One could argue that 3P-KMA has been previously applied at late imaging to healthy liver that showed greater kb values than those of lung tumors; however, we suggest that further studies are warranted to investigate the 3P-KMA efficiency in various tissues showing greater 18F-FDG trapping reversibility than lung cancer lesions [9,10,11, 15]. Furthermore, we also suggest that future studies should compare the performance of the 3P-KMA non-linear fitting with that of a step-wise approach replotting non-linear graphical data with different values of kb in order to recover a linear fitting and hence to obtain Ki [8]. Finally, the current study did not use respiratory gating and, in the case of lesions in the lower lobes, respiratory artifacts may very likely have affected outcomes of both 3P-KMA, Patlak’s analysis, and 5P-KMA and thus might have had an influence on the reported SDs and limits of agreements.

Conclusions

Comparison between 3P-KMA and standard Patlak’s analysis showed that the latter significantly underestimates, on average, the net influx constant (Ki) value in comparison with the former, because it arbitrarily set the release rate constant (kb) to zero: the greater the kb value, the greater the Ki underestimation. This underestimation was not revealed when comparing 5P-KMA and Patlak’s analysis. We suggest that further studies are warranted to investigate 3P-KMA efficiency in various tissues, either physiological or pathological, showing greater 18F-FDG trapping reversibility than lung cancer lesions.