Kinetic Analysis of 3′-Deoxy-3′-Fluorothymidine PET Studies: Validation Studies in Patients with Lung Cancer

Patients

Seventeen patients with 18 biopsy-proven or clinically suspected NSCLCs were prospectively studied. All patients had CT of the chest before FLT PET imaging. These studies were conducted following protocols approved by the University of Washington Human Subjects and Radiation Safety Committees. All patients provided signed informed consent. Data from 10 of these patients have appeared in an earlier report (16).

Surgery and Histopathologic Analysis

Surgical resection of tumor specimens occurred within 1 wk of FLT scanning for all patients except one, who underwent surgical resection at 47 d. One patient had 2 histologically different primary NSCLCs resected, 1 from each lung at 2 different operations. Histologic specimens were available for pathologic and IHC evaluation. All tumor specimens were reviewed by a pathologist who was unaware of the PET results to assess tumor type and differentiation.

A representative formalin-fixed, paraffin-embedded section from each specimen was labeled using monoclonal antibody MIB-1 (Immunotech; 1:100) after microwave antigen retrieval in citrate buffer. Antibody binding was detected using the Vectra Elite kit with FeCl₃ intensification and hematoxylin counterstain. MIB-1 recognizes the Ki-67 antigen, a M_r 345,000 and M_r 395,000 nuclear protein common to proliferating human cells (23). Human lymph node tissue was used as a positive control for proliferating cells. The primary antibody was omitted on sections used as negative controls. All cells with nuclear staining of any intensity were defined as positive. Proliferative activity was defined as the percentage of nuclei stained with MIB-1 per total number of nuclei in the sample. The fraction of labeled tumor cells, defined as the Ki-67 labeling index (Ki-67 LI), was assessed over 4 microscopic fields, 3-mm diameter in the field (estimated 2,000 nuclei), that contained the highest average fraction of labeled cells. The Ki-67 LI is analyzed as a categorical variable due to the method of scoring the fraction of cells stained with the MIB-1 antibody. The categories are determined as 5% intervals with complete staining of every cell as 100%. This Ki-67 IHC method has been used by our group as a tissue correlate to PET measures (16,28).

Radiosynthesis

FLT was prepared according to the method developed by Grierson and Shields (29). The specific activity was >37 GBq/mmol at the time of injection. Before administration of each dose, quality control testing for pH and radiochemical purity was completed. The radiochemical purity was >98% for all FLT injections. All of the doses administered were shown to be free of endotoxin, <0.4 IU/mL (Limulus Amebocyte Lysate; Associates of Cape Cod Inc.). Sterility testing for anaerobic and aerobic bacterial contamination was performed on samples of all of the batch dose after radioactive decay (24 h). FLT was administered by intravenous injection of a 10-mL solution of isotonic saline containing <10% (v/v) ethanol USP. Patient dose was based on the patient’s weight (2.6 MBq/kg), with a 185-MBq maximum.

Imaging Protocol

The PET studies were performed on an Advance PET tomograph (General Electric Medical Systems) providing 35 image planes over a 15-cm axial field of view with a 4.25-mm slice spacing (30). The image study design, injection procedure, and imaging parameters for this group of patients have been described elsewhere (16). Briefly, images were acquired with the following sequence: eight 15-s, four 30-s, six 60-s, two 5-min, and ten 10-min time frames. Images were reconstructed by the method of filtered backprojection using a 10-mm Hanning cutoff filter. Nine patients were dynamically imaged for 120 min, with the remaining patients imaged for 90 min.

Blood Sampling and Metabolite Analysis

Seven patients had either 24 or 30 arterial blood samples assessed for radioactivity throughout the 90 or 120 min of imaging, respectively. The remaining patients had limited venous sampling at 1, 5, 10, 15, 20, 30, 45, 60, 90, and 120 min after injection. Venous samples were used to calibrate image-derived cardiac input functions from left ventricular regions of interest (ROIs) to generate an arterial input function. For each blood sample, 0.1 mL plasma and 0.1 mL whole blood were assayed for radioactivity.

FLT catabolism primarily occurs in the liver to produce a glucuronide conjugate, which is exported to the blood and eventually is cleared by the kidneys (19,20). ¹⁸F-FLT and ¹⁸F-FLT-glucuronide metabolite appears to be the only observed labeled species in blood during the imaging studies. Because FLT has negligible serum protein binding (31), all of the activity associated with FLT in blood was assumed to be available for tissue uptake. We developed a simple solid-phase extraction chromatography (SepPak; Waters Corp.) for the separation of FLT from FLT-glucuronide. An aliquot (0.5 mL) of the plasma from arterial and venous samples was assayed for the relative amount of FLT and FLT-glucuronide. To minimize the number of metabolite determinations, we applied an empiric curve fit similar to approaches we have used in comparable settings (32). The fraction of total activity present as FLT versus the time after injection was fitted to a monoexponential curve with a scaling factor using a nonlinear least-squares regression (NLR) optimization procedure: Eq. 1 The 3 floating variables were s, a scale factor; c, an offset value; and k, a rate parameter. These variables were fit to the FLT assay data for each patient. The curve provides a continuous function describing the fraction of the total plasma activity associated with FLT versus metabolites. The fractional FLT curve was applied to the total blood activity curve to provide a metabolite-corrected FLT input function used for further modeling analysis on individual patient datasets.

PET Image Processing

ROIs for tumor, marrow, and muscle were identified on images summed between 30 and 60 min, an interval when transport of the radiotracer from blood to tissue is predominantly unidirectional. ROI construction with Alice image processing software (Hayden Image Processing Group) was aided by referencing the closely aligned CT images. The ROIs from contiguous slices were combined to create volumes of interest (VOIs) for each tissue type. Tumor and vertebral marrow VOIs were constructed on summed FLT images by creating a perimeter region at 50% of the maximum pixel value of the entire volume. All pixels within this region for each slice were included in the volume. VOIs were applied to the dynamic image set for data extraction. Marrow regions typically covered 15 slices with a total volume of approximately 30 mL. No marrow ROI dimension was <2 cm so that partial-volume correction was not required. Tumor regions were placed on all of the planes containing portions of the lesion that were >50% of the tumor maximum pixel value. Tumors were variable in diameter ranging from 1.5 to 7.7 cm. Those with less than twice the reconstructed resolution of approximately 10 mm in any direction (n = 10), as determined from CT scans, had partial-volume correction applied using a measured recovery coefficient (33) as described in a prior report (16). Muscle VOIs from the arm or back were generated from aligned transmission and CT scans and typically covered 15 slices resulting in a 100- to 200-mL VOI and did not require partial-volume correction.

Quantitative Analysis

Model-independent estimates of FLT uptake were assessed by the SUV and by a modified graphical analysis. The SUV was determined from the activity in each region obtained from the 30- to 60-min summed FLT image (C_i) and normalized to the injected dose and the patient’s weight using the following formula: Eq. 2 The modified graphical determination (34) of FLT flux corrects for blood metabolites, which are confined to the blood pool. This graphical determination of flux may be valid for a short interval after injection, when the approximation of unidirectional transfer of FLT from blood to tissue holds (35,36). The method assumes that phosphorylated FLT nucleotides are retained in the tissue. In the plot of normalized uptake [C_i(t)/C_b (t)] versus normalized time [∫₀^tC_pFLT(τ)dτ/C_b(t)], the slope of the linear portion of the curve estimates the flux of FLT (K_FLT) into the tissue region and occurs after the equilibration phase. The onset of pseudoequilibration of the tissue compartment has been assumed to be 5 times the clearance half-life of the free precursor pool, roughly estimated by 0.693/(k₂ + k₃) (35). From the initial evaluation of FLT kinetic parameters, this time was approximately 15 min, which was used as a lower boundary for graphical analysis. Significant deviation from linearity for graphical analysis was observed after 50 min for most tumor regions. Therefore, 15 and 50 min were chosen as the time boundaries for the linear fit in the graphical analysis.

For each time frame of the dynamic imaging sequence, the average MBq/mL within the VOI was used for compartmental model analysis. The regional VOI activity curves, the metabolite-corrected arterial input curve, and the total arterial activity curve were fitted to the FLT compartmental model (Fig. 1) using the weighted Levenberg–Marquart least-squares minimization algorithm as implemented in a software package designed for PET data analysis (PMOD version 2.5; PMOD group, Zurich, Switzerland (37)). The 2-tissue compartment, 4-rate parameter model (4P) is described in a companion report (27) and is illustrated in Figure 1.

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 1.

Kinetic model of FLT metabolism is comprised of an exchangeable tissue compartment and a compartment of trapped FLT phosphate nucleotides. Four rate constants describe kinetic transfer rates between the 2 compartments. FLTMP = FLT-monophosphate; FLTDP = FLT-diphosphate; FLTTP = FLT-triphosphate; FLT-gluc = FLT-glucuronide; Q_e = exchangeable tissue compartment; Q_m = compartment of trapped FLT phosphorylated nucleotides; C_pFLT = concentration of FLT in arterial plasma; C_met = concentration of metabolites in arterial plasma.

The individual rate constants, K₁, V_d (=K₁/k₂), k₃, and k₄ and the regional V_b, the fraction of vascular activity in the tissue VOI, were allowed to float during the optimization process. Starting values and ranges of the parameters appear in Table 2 of the companion report (27). The residuals were weighted by the inverse variance of the total counting rate in each frame of data, based on the SD of the total uncorrected counts and the duration of the time frame (38). Model parameters were estimated by minimizing the weighted residual sum of the square error (WRRS) between the model solution and the PET measurement. For the datasets > 90 min, model optimization was also performed with the first 90 min of data to generate a complete set of estimates with the same duration. The FLT flux, K_FLT, was estimated by compartmental modeling using parameters derived from fitting the FLT input function and the total blood activity curve to the tissue time–activity curve data: Eq. 3

Once thymidine is incorporated into DNA, it is retained for the duration of the imaging period. However, phosphorylated nucleotides of FLT (FLT-monophosphate, FLT-diphosphate, FLT-triphosphate) can be dephosphorylated back to FLT with the potential to leave the imaging region (21). The loss of image signal through this process, described by k₄, should be identifiable and distinct from other rate constants (27), so that the inclusion of k₄ should improve the fit of the model to the data. We tested this hypothesis with patient datasets from 90 to 120 min of imaging using the corrected Akaike Information Criteria (AIC_c) (39) by comparing a 4P (including k₄) model and a 3-parameter (3P; no k₄) model on identical patient tumor data. The most appropriate model is not necessarily the model producing the smallest WRRS, since adding more parameters generally decreases WRRS. The AIC_c score applies a penalty for additional parameters, which balances the error from fitting the model to the data and minimizing the number of parameters in the model. A lower AIC_c score for a model that includes k₄ would indicate that the 4P model accounts for the data better than a 3P model with a higher AIC_c score.

To test the effect on parameter estimates of using a simpler model (3P) and the shorter imaging period described in one report (17), we analyzed the first 60 min of each dataset using a 3P model (k₄ fixed to 0) and compared the results with the full dataset analyzed using the 4P model. We expected the 3P model to underestimate K_FLT compared with a longer imaging acquisition applying a 4P model that accounts for label loss (36,40). The effect of blood metabolite levels on K_FLT estimation was also examined by fitting the tissue curves with the total blood activity curve.

Statistical Analysis

To assess the relationship between an independent assessment of cellular proliferation and various PET measures of FLT uptake, correlations of Ki-67 LI were made between flux parameters from the 4P model (120 and 90 min of data), the 3P model (60 min of data), graphical analysis (15- to 50-min interval), and the average SUV (30–60 min of data). The correlations between FLT PET uptake values or model estimations and the categoric variable Ki-67 LI were tested using the nonparametric rank correlation by Spearman (ρ). Comparisons between groups of the continuously distributed variables were made using standard parametric statistical tests (2-tailed paired Student t test). Statistical analyses were performed using the statistical software JMP (SAS Institute).

Blood and Tissue Activity Curves

Typical blood and plasma activity curves appear in Figure 2A. There were no differences in FLT concentrations between plasma and whole blood. Arterial and venous blood samples were compared for 6 patients. The average deviation between the total whole blood and plasma activity concentration was 5% ± 8% (mean ± SD, n = 127 samples). There were no significant differences for blood-to-plasma ratios between arterial (mean ± SD, 0.96 ± 0.07) and venous (mean ± SD, 0.97 ± 0.08) samples over the 2-h study period. Metabolite analysis of several blood samples provided the fraction of activity present as FLT. A typical result is shown in Figure 2B.

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 2.

Blood sampling for FLT was acquired through either venous or arterial catheters at various times after injection. (A) Several patients (n = 6) had both blood and plasma activity determined from each sample. (B) Activity in blood after injection of ¹⁸F-FLT is comprised of both FLT and a labeled metabolite (FLT-glucuronide) determined by an independent assay. (C) Combining total activity in blood and the FLT fraction permits generation of an input function for FLT used to quantitate FLT tissue retention. %ID/g = percentage injected dose per gram.

A typical FLT input function (Fig. 2C) is derived from the blood activity in Figure 2A and the metabolite-correction function in Figure 2B. The proportion of FLT and metabolites of FLT in blood showed a predictable time course, with an approximately 5% coefficient of variation (COV) over the set of patient analyses (range, 0.2%–8.1% COV). The average fraction ± SD of the total activity present as labeled metabolites was 25% ± 5% at 60 min (n = 17) and 30% ± 6% at 120 min (n = 9).

An example FLT patient image for 30–60 min after injection and tissue time–activity curves are shown in Figure 3. They demonstrate the anticipated variation in FLT uptake between different tissues. Proliferative normal tissues such as bone marrow showed rapid early uptake and high persistent retention. Muscle showed a much reduced and largely reversible uptake, which mirrored the blood activity a few minutes after injection. Tumor showed uptake between muscle and marrow and often showed a late decline in uptake, similar to the tumor curve in Figure 3A.

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 3.

Three tissue types were investigated for FLT uptake. (A) Example of patient tissue time–activity curves decay corrected to time of injection. (B) Patient PET image acquired 30–60 min after injection of 118.4 MBq FLT. (C) Graphical analysis plot of normalized tissue uptake vs. normalized time for marrow, tumor, and muscle. (D) CT image close to PET image slice provides information for ROI placement and determination of recovery coefficients for tumor regions. Patient’s arms are not in field of view for standard CT protocol.

Parameter Estimates

The mean and range of FLT model parameter estimates for marrow, muscle, and tumor using 120 or 90 min of data are presented in Table 1. The parameter estimates from a subset of the activity curves fit for 60 min of data using the 3P model also appear in Table 1. Model optimization results suggest that FLT is not retained to any significant degree in tissues with low or no proliferation, such as muscle. For tumors imaged for 120 min, we observed a finite rate of late label loss, with an average estimate for k₄ of 0.020 min⁻¹ (range, 0.011–0.050 min⁻¹, n = 10), which was highly correlated (r = 0.91, n = 10) with k₄ estimates using 90 min of data (0.023 min⁻¹, range 0.014–0.052 min⁻¹, n = 10). Consistent with this finding, a characteristic late downward curvature in the graphical analysis relationship was observed for most tumors (for example, Fig. 3C). This was not the case for marrow with little downward curvature and a mean k₄ of 0.006 min⁻¹ (range, 0.001–0.014 min⁻¹).

The inclusion of k₄ in the model for tumors was supported by a lower AIC_c for 15 of the 18 tumor datasets using the 4P model with 90 min of data (Table 2) when compared with a 3P model using the same data. The 3 patients with higher AIC_c values for the 4P model had estimated tumor k₄ values close to zero (0.003, 0.007, and 0.007), much lower than the mean k₄ for the group of 0.017 min⁻¹ (n = 18).

The model analysis in our companion report (27) revealed that individual parameters covaried significantly but that 2 parameters robustly and independently estimated transport (K₁) and overall flux (K_FLT). Additionally, K₁ is primarily determined in the first few minutes after injection (27). The estimation of initial transport is expected to be similar between the 3P and 4P models. We therefore compared K_FLT for the different analytic methods to assess the effect of model choice on parameter estimates. Flux values (K_FLT) obtained from the 4P model with 120 min of data (mean, 0.064 mL/min/g; range, 0.014–0.099 mL/min/g; n = 10) correlated with K_FLT values estimated from graphical analysis (mean, 0.047 mL/min/g; range, 0.011–0.076 mL/min/g; r = 0.86), a 4P model using only 60 min of data (mean, 0.058 mL/min/g; range, 0.005–0.101 mL/min/g; r = 0.87), and K_FLT estimates from a 3P model using only 60 min of data (mean, 0.047 mL/min/g; range, 0.009–0.071 mL/min/g; r = 0.94) and are presented in Table 3. The K_FLT estimates from the 4P model with 120 min of data were compared with average SUV values 30–60 min after injection and showed a poor correlation (mean, 4.03; range, 0.99–7.55; r = 0.62; n = 10).

However, K_FLT estimates from the 90-min 4P model were consistently higher compared with the 3P model estimates (mean, 26%; range, 10%–46%), as illustrated in Figure 4 and Table 4. The 90-min 4P model estimates of K_FLT were also higher than estimates using a 4P model with 60 min of data (mean, 16%; range, −19% to 70%) and the graphical approach (mean, 30%; range, 1%–67%). In a separate analysis of the effect of blood metabolites on parameter estimates, the assumption that the total blood activity was entirely FLT would lead to an underestimate of K_FLT by as much as 47%.

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 4.

Estimated K_FLT value from 1 h of imaging and blood data using 3P model show high correlation with K_FLT estimates using 4P model (r = 0.95; n = 18). The 60-min determination that ignores k₄ consistently underestimated the K_FLT value determined using more data and the 4P model.

Correlation with Ki-67 LI

Eighteen tumor specimens were collected by biopsy after FLT PET, and sections were analyzed by Ki-67 IHC. K_FLT from the 4P model (either 120- or 90-min sets) versus Ki-67 LI showed a high degree of correlation (Spearman ρ = 0.92; P < 0.0007) (Fig. 5). A similar comparison, but using a 3P model with 60 min of data, had a Spearman correlation of ρ = 0.87 (P < 0.0001; n = 18), and a 4P model with 60 min of data had a correlation of ρ = 0.71 (P < 0.0011; n = 18) (Table 5).

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 5.

Ki-67 LI determination from tissue biopsy specimens allows for independent assessment of cellular proliferation. Plot of FLT flux determined from 4P model vs. Ki-67 LI shows high degree of correlation (ρ = 0.92), suggesting that use of FLT is appropriate for noninvasive estimation of cellular proliferation.