Abstract
The most accurate way to estimate the glucose metabolic rate (or its influx constant) from 18F-FDG PET is to perform a full kinetic analysis (or its simplified Patlak version), requiring dynamic imaging and the knowledge of arterial activity as a function of time. To avoid invasive arterial blood sampling, a simplified kinetic analysis (SKA) has been proposed, based on blood curves measured from a control group. Here, we extend the SKA by allowing for a greater variety of arterial input function (A(t)) curves among patients than in the original SKA and by accounting for unmetabolized 18F-FDG in the tumor. Methods: Ten A(t)s measured in patients were analyzed using a principal-component analysis to derive 2 principal components describing most of the variability of the A(t). The mean distribution volume of 18F-FDG in tumors for these patients was used to estimate the corresponding quantity in other patients. In subsequent patient studies, the A(t) was described as a linear combination of the 2 principal components, for which the 2 scaling factors were obtained from an early and a late venous sample drawn for the patient. The original and extended SKA (ESKA) were assessed using fifty-seven 18F-FDG PET scans with various tumor types and locations and using different injection and acquisition protocols, with the Ki derived from Patlak analysis as a reference. Results: ESKA improved the accuracy or precision of the input function (area under the blood curve) for all protocols examined. The mean errors (±SD) in Ki estimates were −12% ± 33% for SKA and −7% ± 22% for ESKA for a 20-s injection protocol with a 55-min postinjection PET scan, 20% ± 42% for SKA and 1% ± 29% for ESKA (P < 0.05) for a 120-s injection protocol with a 55-min postinjection PET scan, and −37% ± 19% for SKA and −4% ± 6% for ESKA (P < 0.05) for a 20-s injection protocol with a 120-min postinjection PET scan. Changes in Ki between the 2 PET scans in the same patients also tended to be estimated more accurately and more precisely with ESKA than with SKA. Conclusion: ESKA, compared with SKA, significantly improved the accuracy and precision of Ki estimates in 18F-FDG PET. ESKA is more robust than SKA with respect to various injection and acquisition protocols.
PET using 18F-FDG is useful for grading tumors and assessing therapy or disease progression (1,2). 18F-FDG uptake is often characterized by calculating the standardized uptake value (SUV) from late static imaging (typically 45–60 min after injection). However, SUV is sensitive to body composition (3), change in uptake over time (4), and blood glucose concentration (5), which compromise the use of SUVs for inter- or intrapatient comparison (6). Several normalization schemes have been proposed to reduce this sensitivity (7) but do not account for variations in 18F-FDG pharmacokinetics among individual patients, and these methods do not differentiate metabolized from unmetabolized 18F-FDG within the tumor. The most accurate way to characterize the glucose metabolic rate is to perform a full kinetic analysis (or its simplified Patlak version (8)) to estimate the influx constant (Ki) of 18F-FDG in the tumor. Computing Ki, however, requires dynamic PET to assess the tumor uptake kinetics (tumor uptake at time t [T(t)]) and repeated blood sampling or dynamic imaging over the heart or aorta to measure the arterial input function A(t).
As a compromise between SUV and Patlak analysis, the so-called simplified kinetic analysis (SKA) has been proposed (7). In this method, the shape of the patient's A(t) is assumed to follow a mathematic model derived from a group of subjects. For any patient, the A(t) magnitude is obtained by scaling the A(t) model to fit 1 late venous blood sample. A drawback of this approach is that it assumes that a single function can model every individual patient's A(t) after appropriate scaling (1,7,9). In addition, SKA does not differentiate metabolized from unmetabolized 18F-FDG in the tumor. To overcome the SKA limitations, we propose an extended SKA method (ESKA), which allows for a greater variety of A(t) curves among patients and accounts for unmetabolized 18F-FDG in the tumor. We compare the performance of SKA and ESKA, using Patlak analysis as a reference.
MATERIALS AND METHODS
Patlak Analysis
In the Patlak method (8), the Ki of 18F-FDG in the tumor is given by:
SKA
The SKA method initially proposed by Hunter et al. (7) requires 1 static acquisition and 1 venous blood sample collected midway through the static acquisition, approximately 55 min after injection of 18F-FDG. In this approach, A(t) is modeled using a triexponential function:
In our study, we implemented 2 versions of SKA. The first was SKApub, for which we used the bi parameters provided in the papers by Hunter et al. (7), Graham et al. (10), and Hoekstra et al. (11), given they were supposed to stay in a narrow range for all patients. The second version was SKAopt, for which we used our own patient studies to derive the bi parameters, by considering the decreasing part of the true A(t).
In both approaches, A1 = A2 in Equation 2 and was computed as proposed in the papers by Hunter et al. (7), Graham et al. (10), and Hoekstra et al. (11). The lean body mass equations were 1.07 × weight – 148 × (weight/height2) for men and 1.1 × weight – 120 × (weight/height2) for women (12). A3 was computed for each patient, using a late A(tlate) value, where tlate corresponded to the end of the scan.
ESKA
Similar to SKA, ESKA is based on an A(t) model, derived from a training subset of randomly chosen patients for whom the fully sampled A(t) was available.
In ESKA, estimating the A(t) relies on the following 3 steps (Fig. 1):
Resampling all patients’ A(t) from the training set to the same time sampling.
Analyzing the set of resampled A(t) using a principal-component analysis (PCA). We empirically found that the first 2 principal components (APCA-1(t) and APCA-2(t)) associated with the largest 2 eigenvalues were sufficient to accurately describe any individual A(t):
Taking venous blood samples for any patient not in the training set; the a1 and a2 coefficients are obtained from 2 samples, AESKA(t1) and AESKA(t2). AESKA(t1) was taken early after injection but late enough to ensure arteriovenous activity equilibrium. AESKA(t2) was taken at the end of the scan. These 2 values, along with the APCA(t1) and APCA(t2) values, yielded a system of 2 equations with 2 unknowns, from which a1 and a2 could be derived.
Three steps of ESKA approach for estimation of A(t).
Unlike SKA, ESKA accounts for the unmetabolized 18F-FDG in the tumor. Indeed, a mean distribution volume (Umean) was estimated from the group of patients in the training set (1). For each individual patient, the ESKA Ki index, which estimates Ki when using the ESKA method, was then given by:
Patients
Forty-three patients were considered and divided into 3 groups. Each group corresponded to a particular injection and acquisition protocol, to assess the robustness of the method for various injection durations and delays between injection and scanning. Nineteen patients were studied multiple times (baseline followed by 1 or 2 follow-up studies), resulting in 70 studies, each with the complete A(t) available for reference (Table 1). Pre- and posttherapy scans were combined to compare the various kinetic analysis methods.
Patient Population
Results from the first 2 groups assessed the robustness of the methods to the injection protocol:
The 120-s/55-mTot group (33 studies) included 16 patients (47.1 ± 13.7 y) with renal cell cancer receiving vascular endothelial growth factor antibody. Each patient underwent a 120-s 18F-FDG injection protocol using a constant-infusion pump. These patients were scanned dynamically for about 55 min after injection.
The 20-s/55-mTot group (37 studies) included 27 patients (47.6 ± 17.3 y) who underwent a 20-s 18F-FDG injection protocol using a constant-infusion pump. These patients were enrolled in 3 different protocols: 13 breast cancer patients were receiving anti–epidermal growth factor receptor therapy, 7 patients with various tumor locations were receiving various types of antineoplastic therapy, and 7 patients were admitted for diagnosis of lymph node lesions. These patients were scanned dynamically for approximately 55 min after injection.
To investigate the robustness of the methods regarding the delay between injection and scanning, another group was derived from the 20-s/55-m group, considering delayed data:
The 20-s/120-mTot group (7 studies) included the 7 patients from the 20-s/55-mTot group (25.7 ± 13.3 y) presenting lymph node inflammation for whom we had a delayed whole-body scan available. This delayed whole-body scan allowed assessment of the tumor at 1 additional late time point, approximately 90–150 min after injection depending on the lesion localization. In this group, we considered tumors only with stable Patlak Ki—that is, with Ki that did not vary by more than 15% between the injection groups at 55 and 120 min.
Patient Training and Testing Groups.
To derive the A(t) models (SKA parameter bi and the 2 principal components APCA-1(t) and APCA-2(t) needed for ESKA), some of the patients from the 120-s/55-mTot and 20-s/55-mTot groups were removed to be used as a training set. The training subjects used for computing these parameters or principal components were not further included in the analyses. The same training subjects were used for SKA and ESKA. Thus, in the testing set, 10 subjects were in the 120-s/55-m, 21 in the 20-s/55-m, and 7 in the 20-s/120-m groups (Table 1).
We also used pairs of successive PET/CT scans acquired in the same subjects to study the accuracy and precision in measuring changes in Ki between 2 PET/CT scans (ΔKi). After removal of the cases used in the training set, we had 13 tumors for which ΔKi between 2 scans was available in the 20-s/55-m group and 31 tumors in the 120-s/55-m group, with an average time lapse between the 2 scans of 34.6 ± 14.8 and 42.5 ± 10.0 d, respectively (none in the 20-s/120-m group).
Lesions Considered in Study.
The field of view explored on the first PET scan was chosen to include lesions greater than 2 cm in their longest dimension based on CT or MRI data. All additional lesions in the field of view at the time of the first dynamic PET scan were also considered. Overall, a total of 124 lesions were investigated.
Patient Dose.
For each patient, approximately 370 MBq (379 ± 61 MBq) of 18F-FDG were injected, except for the 7 patients with lymphatic lesions (560.1 ± 14.8 MBq for the 4 adults and 201.5 ± 75.4 MBq for the 3 children).
Data Acquisition
For all patients, 18F-FDG PET studies were acquired dynamically, starting at injection time, on an Advance PET (GE Healthcare) scanner in 2-dimensional mode, producing 35 slices over a 15-cm axial field of view.
For the 120-s injection protocol (the 120-s/55-m group), the time per frame was 30 s (0–4 min after injection), 3 min (4–40 min after injection), and 5 min thereafter, for an average of 51.3 ± 2.2 min scan time.
For the 20-s injection protocol (the 20-s/55-m and 20-s/120-m groups), the time per frame was 5 s (0–1 min after injection), 15 s (1–2 min after injection), 30 s (2–7 min after injection), 3 min (7–25 min after injection), and 5 min thereafter, for an average of 54.7 ± 1.1 min scan time.
When tumors were not at the level of the heart (17 studies in total), the scan was started over the heart for a period of approximately 25 min and then shifted to the tumor level for the remaining dynamic acquisition time.
A static image at the tumor level was systematically computed from the dynamic acquisition, summing the last 3 acquisition time points (∼45–60 min after injection).
In addition to the dynamic acquisition, the 7 patients in the 20-s/120-m group underwent a whole-body static acquisition at the end of the dynamic acquisition, on which the node and heart uptake was seen at the late time point (between 90 and 150 min, depending on the number of steps before reaching the tumor location, step duration, scan starting time). The time associated with the node (or the heart) level was computed from the whole-body starting time and the slice including the node (or the heart).
Data Processing
Image Reconstruction.
Images were reconstructed at 2 mm/pixel into a 256 × 256 matrix using ordered-subset expectation maximization (4 iterations, 28 subsets, 5-mm gaussian postreconstruction filtering) including attenuation, scatter, random, dead time, and decay corrections. The reconstructed images had a transaxial and axial resolution of approximately 7 mm.
Input Function.
The A(t)s were needed for Patlak analysis, which was used as the reference method against which we compared SKA and ESKA. If the tumors were at the level of the heart, A(t) was directly derived from the dynamic images by placing a volume of interest manually in the left atrial cavity (as visualized from the sum of the early arterial phase dynamic images), because this structure is less affected by activity spillover from atrial muscle than the left ventricle cavity, which can be strongly affected by spillover from the myocardium. When the tumor was not at the level of the heart, A(t) was also estimated from the images for the first 25 min (because the first 25 min of data were always acquired at the level of the heart). Then, when the scanner was moved to the level of the tumor, the later A(t) time points were obtained from blood samples taken approximately every 5 min thereafter.
In the Patlak analysis, the data from 25 min after injection to the end of acquisitions were used to obtain the slope Ki and the distribution volume U.
In SKA and ESKA, the input function was estimated using tlate (∼55 min for the 20-s/55-m and 120-s/55-m groups and ∼120 min for the 20-s/120-m group).
Volumes of Interest.
Volumes of interest were drawn over the tumors on each patient from the late tumor static scan using an automatic 3-dimensional region-growing program based on a threshold in uptake (i.e., a percentage of maximum voxel intensity) as implemented in the MedX software (Sensors System). These volumes of interest were used for calculation of the SKA Ki and ESKA Ki values and the Patlak Ki. For the whole-body images, we computed the tumor volumes of interest with the same 3-dimensional region-growing program, using the same tumor uptake threshold as when using dynamic images.
Data Analysis
Performances of SKApub, SKAopt, and ESKA were assessed by comparing the estimated A(t) with the PET image–derived A(t). A percentage error between the estimated area under the curve (AUC) and the AUC derived from the continually sampled A(t) was calculated.
SKApub Ki, SKAopt Ki, and ESKA Ki were compared with our gold standard Patlak Ki using correlation analyses and percentage errors. Estimated ΔKi based on estimated Ki from SKApub, SKAopt, and ESKA were also compared with ΔKi based on Patlak Ki, to study the accuracy and precision in measuring changes in Ki.
Averaged percentage errors, with the percentage error defined by 100 × (estimated_value – reference_value)/reference_value, and associated variance were compared using the nonparametric Wilcoxon matched-pairs signed-rank test and Friedman test, respectively, with an α = 0.05 level of significance. Statistical difference of the nonparametric Spearman correlation coefficients was computed using the Fisher r-to-Z transformation.
RESULTS
We first evaluated the methods for the 120-s/55-m and 20-s/55-m groups, corresponding to acquisitions of about 55 min obtained using two 18F-FDG infusion protocols. We then tested the robustness of the 3 methods for the delayed imaging protocol.
A(t) and AUC Estimates
Figure 2 shows the A(t) (first 10 min only), obtained using SKApub, SKAopt, and ESKA, compared with the reference A(t) for the training and test subsets of the 20-s/55-m group. Large differences between estimated and true A(t) can be seen with SKApub. This result was expected because SKA was not meant to reproduce the shape of the A(t)s at early times. However, a good agreement between the estimated A(t) and the reference A(t) was observed with both SKAopt and ESKA.
(Left) Comparison of original A(t) (first 10 min) with SKA method of Hunter et al. (7), using both published and optimized parameter values to compute ASKA(t). (Right) For same patient studies, comparison of AESKA(t) with original resampled A(t).
Table 2 gives the percentage errors in AUC estimated for SKApub, SKAopt, and ESKA, compared with the reference AUC and the correlation coefficients between reference and estimated AUC. With ESKA, the maximum error in AUC estimate for the 120-s/55-m and 20-s/55-m groups was 8.4%, and the mean error was −0.3% ± 4.3% (computing a combined error for the 2 groups). For these same groups, the maximum error reached 193.8% and 56.7% for SKApub and SKAopt, respectively, and the combined-group mean errors were 52.9% ± 45.2% and 9.1% ± 15.0%, respectively. The correlation coefficients for ESKA were always significantly higher than the ones obtained for the SKA approaches.
Mean Percentage Differences (±SD) Between Estimated and True AUC and Spearman Correlation for SKA and ESKA
The trends for the 20-s/120-m group were the same as those observed for the 120-s/55-m and 20-s/55-m groups. SKApub resulted in the largest AUC overestimation for all groups. The decrease in the mean percentage error between SKApub and SKAopt and the decrease in the mean percentage error between SKAopt and ESKA were statistically significant. The correlation coefficient was significantly higher for ESKA than for SKAopt and SKApub, except for the 20-s/120-m group (possibly because of the low number of subjects in that group). For all methods, except again for the 20-s/120-m group, the magnitude of the error in AUC estimate was not correlated with the magnitude of the reference AUC (results not shown).
Ki Estimates
Table 3 presents the mean percentage error in the estimation of the Patlak Ki for the 3 approaches and the r2 values between estimated and reference Ki. Although the AUC estimates were more accurate for SKAopt than for SKApub (Table 2), the Ki values from SKAopt were significantly less accurate than the Ki values from SKApub. Ki estimates were the most accurate with ESKA, although the percentage errors were highly variable. To test whether the large SD of errors for ESKA was due to using a single Umean to estimate the 18F-FDG distribution volume, we used the true tumor's distribution volume (Utrue) of 18F-FDG computed from the Patlak analysis with the known T(t) and A(t) (results not shown). When accounting for Utrue in ESKA, the SD associated with the mean percentage error in Ki estimates was strongly reduced (mean errors of −6.4% ± 15.2% and 0.9% ± 4.9% for the 120-s/55-m and 20-s/55-m groups, respectively). The impact on the error induced by the use of a unique Umean was statistically significant (P < 0.05) only for the 20-s/55-m group. Utrue values ranged from 0.1 to 0.9 vs. 0.5 for the Umean value.
In the 20-s/120-m group, ESKA also resulted in a statistically better estimate of Ki than did SKA (P < 0.05), associated with a lower SD (P < 0.05 for SKApub vs. ESKA and for SKAopt vs. ESKA).
Mean Percentage Differences (±SD) Between Estimated Ki and Patlak Ki and Spearman Correlation for SKA and ESKA
For all methods, the magnitude of the error in Ki estimate was not correlated with the magnitude of the reference Ki (results not shown).
The correlation coefficients between the reference and estimated Ki with the SKA methods were similar to or significantly lower than those with the ESKA methods. The ESKA Ki values were more strongly correlated with the Patlak Ki values for the delayed acquisitions than for the 55-min acquisitions (P < 0.05 for the 120-s/55-m vs. 20-s/120-m groups and for the 20-s/55-m vs. 20-s/120-m groups).
ΔKi Estimates
Table 4 shows the relative errors (estimated_ΔKi – Patlak_ΔKi)/Patlak_ΔKi when ΔKi was calculated on the basis of SKApub, SKAopt, and ESKA Ki estimates for the 20-s/55-m and 120-s/55-m groups (no data available for the 20-s/120-m group). The smallest mean errors in ΔKi were obtained with ESKA, followed by SKAopt and then SKApub. The errors were always highly variable (SD > 100%), but the smallest variability was observed with ESKA. The magnitude of the error in ΔKi estimate was not correlated with the magnitude of the reference ΔKi (results not shown).
Percentage Errors in Estimation of Change in Ki Between 2 Successive PET/CT Scans When Using SKApub, SKAopt, or ESKA Ki Estimates
DISCUSSION
In 18F-FDG PET, the glucose metabolic rate is related to the net rate of 18F-FDG influx Ki, which can be estimated using a full kinetic analysis and a measurement of the A(t) or using a Patlak analysis (8). However, full kinetic analysis and Patlak analysis are rarely used in clinical practice because of the time-consuming, full dynamic acquisition protocol they involve. In clinical practice, glucose metabolic rate is usually replaced with the SUV, because use of SUVs requires only 1 late image acquisition. Freedman et al. (1) showed that different conclusions could be reached regarding tumor follow-up depending on whether SUV or Patlak analysis was used.
SKAs differ in the method used to estimate the amount of 18F-FDG delivered to the tumor. The SKA proposed by Hunter et al. (7) uses a single parametric expression for A(t) for all patients and a single blood sample to scale the curve for each subject. Adaptation to different infusion protocols or to different patient populations is only through adjustment of the exponential parameters (as in our SKAopt) and by scaling the resulting A(t) using the single blood sample, to correctly estimate the AUC. The A(t) shape actually depends on the injection rate and on the patient's metabolic status. In SKA, this variability is considered using only the scaling factor associated with the third exponential function of the model (Eq. 2). The injection phase is thus neglected, unlike in our ESKA approach. Nor does SKA attempt to account for unmetabolized 18F-FDG in tumor. Sundaram et al. (9) proposed a hybrid model between the SKA method proposed by Hunter et al. (7) and Patlak analysis, using multiple blood samples to account for unmetabolized 18F-FDG. However, only 1 blood sample (40 min after injection) was used to scale each individual patient's A(t). Another approach has been proposed to estimate A(t) using a triexponential function (13) for Na18F-PET data. The authors (13) reported a good correlation between true and estimated A(t) but used the same population of patients to determine the model parameters and validate the model.
Other groups used image-derived input functions without scaling (14,15), an approach limited by the PET image partial-volume effect and radiolabel uptake in tissue surrounding the region of interest.
In this paper, we first considered an approach that optimizes the model parameters for each injection protocol—an alternative version of the SKA method proposed by Hunter et al. (7). Our approach might be more appropriate than the original SKA method when the injection protocol varies. We also derived an alternative model for A(t), with 2 principal components derived from the PCA of a subset of fully sampled A(t) measured using a specific injection protocol. We found (results not shown) that considering more principal components increased variability without reducing error in AUC estimates. Naganawa et al. (16) used independent component analysis to estimate the A(t), along with the tissue time–activity curve, directly from dynamic brain PET series. To deal with the nonunique solution of independent component analysis, these authors introduced anatomic constraints specific to cerebral physiology, which are difficult to adapt in an oncology situation. Their method has the advantage of being independent of the injection protocol and image time sampling and not requiring any manual delineation of the arterial region of interest (16). However, an arterial blood sample was required to scale the estimated input function.
ESKA requires 2 venous blood samples rather than the single sample required by SKA. We do not consider this requirement a significant drawback, because once the patient has a venous line in place to draw the 1 sample required by SKA, it is simple to acquire an additional sample later. To improve the robustness of the fit, 1 early and 1 late sample are required. The first blood sample should be drawn as soon as possible after equilibration between arterial and venous blood has been reached, to accurately estimate the early part of the A(t). From visual examination of the A(t) of the training sets, we chose to draw our first sample at approximately 15 min after injection for the 120-s injection protocol and approximately 10 min after injection for the 20-s injection protocol. Neither the SKA nor the ESKA methods require dynamic imaging. Only a single, late static image is necessary.
Another advantage of the ESKA method is that unlike SKA, ESKA attempts to account for the presence of unmetabolized 18F-FDG when deriving the Ki, based on the average distribution volume of 18F-FDG observed in the training set.
A(t) and AUC Estimates
As illustrated in Figure 2, the A(t) curves recovered using ESKA agree well with the original A(t). ESKA yielded both a better estimate of the AUC and a better correlation with the true AUC than did SKA, for all groups (Table 2).
Ki Estimates
For the 120-s/55-m group, our correlations between SKA Ki and Patlak Ki are consistent with values reported elsewhere in the literature (7,9,10). We found an r2 coefficient between SKApub and true Ki of 0.91, whereas previous correlations of 0.89 (10), 0.92 (9), and 0.98 (7) have been reported. However, this high correlation is accompanied by an overestimation of the SKApub Ki indices similar to that reported by Hunter et al. (7) (although Sundaram et al. (9) reported a 20% underestimation of Ki when using the approach of Hunter et al.).
For the 20-s/55-m group, true Ki and SKApub Ki agreed less well (r2 = 0.59), although there was only a small underestimation of the SKApub Ki index, underlining the variability in SKA performance as a function of the injection protocol. SKApub Ki correlated slightly better with Patlak Ki for 120-min acquisitions than for 55-min acquisitions (Table 3).
There was excellent correlation between Patlak Ki and ESKA Ki for both 20-s and 120-s injections (r2 ≥ 0.92). The mean error in Ki with ESKA was also small for both injection protocols, compared with SKAopt (Table 3). To clarify whether the improvement in Ki estimates observed with ESKA was mostly due to better A(t) estimate or to accounting for unmetabolized 18F-FDG, we calculated the mean error in Ki estimates using SKApub and SKAopt with U set to Umean (as in ESKA) and using ESKA with Umean = 0 (Table 3). Table 3 demonstrates that the 2 original features of ESKA (original A(t) estimate and accounting for Umean) are needed to make ESKA accurate, whatever the injection protocol and acquisition time, and that SKA cannot perform as well as ESKA, even when accounting for unmetabolized 18F-FDG. With ESKA, a residual error was still observed due to the use of a mean Umean value. Error was reduced in the 120-min acquisitions, compared with 55-min acquisitions, presumably because of the decrease of unmetabolized 18F-FDG with time. Also, the SD associated with the mean error was reduced to 6.2% (20-s/120-m group), compared with 28.8% and 22.4% for the 120-s/55-m and 20-s/55-m groups, respectively (P < 0.05).
Compared with ESKA, the larger SD in the SKA methods was due to not accounting for unmetabolized 18F-FDG, as demonstrated in Table 3, in which the SD systematically decreased when Umean was included in the SKA methods.
SKApub Versus SKAopt
In all our patient groups, the AUC was overestimated using SKApub. The results in Figure 2 show the need to adapt the parameters to the injection protocol if the objective is to estimate the A(t) shape accurately. However, even when optimizing the SKA parameters for our injection protocols, SKA still resulted in an AUC overestimation, probably because a single AUC shape, even scaled, cannot be appropriate for every patient. Using SKAopt instead of SKApub left the correlation with true Ki unchanged or improved the correlation, but it did not necessarily improve the Ki estimates, because with SKApub the underestimation of Ki due to the overestimation of AUC is partially compensated by not accounting for unmetabolized 18F-FDG (i.e., overestimating Ki).
ΔKi Estimates
Results in Table 4 show that ESKA tends to estimate ΔKi more accurately and more precisely than SKA, demonstrating that the reduction of bias in Ki estimates obtained using ESKA is not at the expense of increased variability. These results are consistent with the correlation results shown in Table 3 and suggest that ESKA might be preferable to SKA when using Ki for patient monitoring.
Limitations of ESKA
A source of A(t) variability is the time lag between the start of the scanning and the arrival of activity in the field of view. To remove such variability and correctly align the A(t) peak time in ESKA, we aligned the original A(t)s during the resampling step so that each A(t) peaked at the same time point. This realignment can be done a posteriori, as we did here, or by synchronizing the scanning start time with the detection of activity in the field of view.
In ESKA, the unmetabolized 18F-FDG in the tumor is accounted for using the Umean obtained from the training set from each population (Eq. 5). However, only a true kinetic analysis could determine the exact distribution volume. Using only a mean value may cause variability in Ki, as previously reported (1). Variability in tumor distribution volume likely reflects real variations in tumor viable cell density, edema, and other factors that are hard to predict on the basis of other observables—that is, tumor or background uptake, input function, patient weight, and Ki estimate. Note that delayed imaging protocols (e.g., 120 min after injection or longer) have recently proven of clinical interest (17–19). Because unmetabolized 18F-FDG is usually quite low in these cases, the errors in estimating unmetabolized 18F-FDG by ESKA should be less important.
CONCLUSION
By accounting for variability in the shape of A(t) using 2 venous blood samples and partially accounting for unmetabolized 18F-FDG in the tumors, the ESKA approach proposed here yields significantly more accurate and more precise estimates of the net 18F-FDG Ki than those achieved with the conventional SKA. These improvements also translate into more accurate estimates of changes in Ki over time.
DISCLOSURE STATEMENT
The costs of publication of this article were defrayed in part by the payment of page charges. Therefore, and solely to indicate this fact, this article is hereby marked “advertisement” in accordance with 18 USC section 1734.
- © 2011 by Society of Nuclear Medicine
REFERENCES
- Received for publication May 27, 2010.
- Accepted for publication January 7, 2011.
Jump to section
Related Articles
Cited By...
- No citing articles found.