Abstract
18F-fluoroazomycinarabinoside (18F-FAZA) is a hypoxia-specific PET tracer. In future clinical applications of hypoxia imaging, such as early response monitoring or radiation therapy dose painting, accurate quantification of tracer uptake at the voxel level will be required. The aim of the present study was to assess the validity of parametric methods for the quantification of 18F-FAZA studies. Methods: Dynamic 70-min 18F-FAZA scans were obtained from 9 non–small cell lung cancer patients. Arterial blood samples, collected at 7 time points, were used for preprocessing an image-derived input function derived from volumes of interest (VOIs) defined within the ascending aorta. Time–activity curves derived from various tumor VOIs were fitted using nonlinear regression analysis (NLR) to a reversible 2-tissue-compartment model, providing volumes of distribution (VT) as an outcome measure. Next, parametric images were generated by use of both Logan graphic analysis with various linear regression start times and spectral analysis with multiple sets of basis functions. The previously defined tumor VOIs were projected onto these parametric images, and the resulting VT were compared with those obtained from NLR. In addition, the results were compared with tumor-to-blood ratios (SUVr), which are more easily obtainable. Results: The highest correlations and correspondence with NLR-derived VT were found for Logan graphic analysis with a start time of 30 min after injection (R2, 0.88; intraclass correlation coefficient [ICC], 0.93) and for spectral analysis–derived VT with a set of 30 basis functions with exponents ranging from 0.0175 to 1.9 (R2, 0.79; ICC, 0.81). SUVr yielded similar correlations but showed significant bias at high VT (R2, 0.85; ICC, 0.80). Conclusion: Both Logan graphic analysis and spectral analysis yielded VT that showed high correlations with nonlinear regression analysis–derived VT. SUVr showed bias at high VT.
Under hypoxic conditions, tumor cells are more resistant to radiation therapy (1) and their migratory and invasive potentials are increased (2). Imaging of hypoxia may aid in determining patient prognosis and selecting appropriate therapeutic strategies (3–6). In addition, mapping of hypoxic tumor areas may enable specific targeting of hypoxic tumor tissue with intensified radiation therapy (7–9).
18F-fluoroazomycinarabinoside (18F-FAZA) is a promising hypoxia-specific PET tracer (10–15) that is based on a nitroimidazole group and that undergoes electron reduction when it enters cells. Under normoxic conditions, this process is followed by reoxygenation and transport from the cell. In contrast, under hypoxic conditions, the tracer is retained because it binds to macromolecules within cells (16), leading to increased PET signal in areas containing hypoxic cells. However, the signal depends not only on the oxygenation status of the tissue under investigation but also on delivery (plasma concentration, blood flow, and extraction) and nonspecific uptake in tissue. For quantification of the degree of hypoxia, a tracer kinetic model that describes tracer kinetics in relation to delivery through the arterial circulation is needed.
In a previous study, a 2-tissue reversible model with a blood volume parameter (2T4k+VB) was identified as the optimal model for describing 18F-FAZA kinetics (17). Through nonlinear regression analysis (NLR), this model provided excellent fits to measured time–activity curves. Unfortunately, this method is markedly sensitive to noise and therefore is not suitable for analysis at the voxel level. However, the latter type of analysis may be useful given the expected heterogeneity in the distribution of hypoxia within a tumor.
Several methods for linearization have been proposed to reduce the sensitivity of the analysis to noise and enable voxel-level analysis, that is, to enable the generation of parametric images; these methods include spectral analysis (SA) (18), Patlak graphic analysis (19), and Logan graphic analysis (LGA) (20). The purpose of the present study was to assess the validity of these parametric methods for analyzing 18F-FAZA data by comparing the results of parametric analyses and NLR in patients with non–small cell lung cancer. In addition, the results derived from parametric methods were compared with those derived from simplified methods because the latter methods would require a less elaborate imaging protocol.
MATERIALS AND METHODS
Inclusion Criteria
Patients aged 18 y and older were asked to participate in the study if they met the following criteria: histologically or cytologically proven non–small cell lung cancer with a lesion measuring at least 2.5 cm in diameter and located within 15 cm of the aortic arch and a hemoglobin level of at least 7.0 mmol/L. Exclusion criteria were chemotherapy or radiation therapy received within 6 mo before the PET study, pregnancy, and inability to remain supine for 100 min. The study was approved by the Medical Ethics Review Committee of the VU University Medical Center. Before inclusion, each patient signed a written protocol-specific informed consent form.
Data Acquisition
Scanning Protocol
Patients were scanned on a Gemini TF-64 PET/CT scanner (Philips). Each patient underwent a low-dose CT scan (50 mAs; 120 kVp), followed by a 70-min dynamic emission scan. The start time of the latter scan coincided with the beginning of an intravenous bolus injection of 168 ± 39 (mean ± SD) MBq of 18F-FAZA with a specific activity of 143 ± 70 GBq⋅μmol−1; the injection was administered and flushed with saline (5 mL at 0.8 mLs−1 and then 35 mL at 2 mLs−1) using an automated pump system. PET data were normalized and corrected for dead time, randoms, scatter, and decay. Attenuation correction was based on the CT scan, and a 3-dimensional row-action maximum-likelihood reconstruction algorithm was used to reconstruct the data into 37 frames (1 × 10, 8 × 5, 4 × 10, 2 × 15, 3 × 20, 2 × 30, 6 × 60, 4 × 150, 4 × 300, and 3 × 600 s) with a matrix size of 144 × 144 × 45 voxels (voxel size, 4 × 4 × 4 mm). All data were reconstructed in accordance with recently published guidelines for quantitative 18F-FDG PET studies (21,22).
Blood Sampling
At set time points (5, 10, 20, 30, 50, 60, and 70 min after injection), arterial blood samples were manually collected from a cannula inserted into the radial artery. These samples were analyzed for plasma–to–whole-blood ratios and for plasma fractions of parent 18F-FAZA.
Kinetic Analysis
Multiple volumes of interest (VOIs) were defined within tumor tissue. Time–activity curves derived from these VOIs were fitted using NLR. The VOIs were also projected onto parametric images derived from dynamic PET data with LGA (20) and SA (18) and onto simplified parameter images derived from static PET imaging. The resulting VOI-based parameter values were compared.
NLR (Gold Standard)
The governing equations for the 2T4k+VB model are as follows (23):Eq. 1Eq. 2andEq. 3where C(t) represents the activity concentrations over time in tissue (CT), in the first compartment (free tracer and nonspecifically bound tracer in tissue; CND), in the second compartment (specifically bound tracer in tissue; CS), and in plasma after correction for metabolites (CP). Because each VOI encompasses some blood volume, blood volume correction was applied with the following equation:Eq. 4where VB is the blood volume parameter, CWB(t) is the whole-blood time–activity curve, and CPET(t) is the tissue time–activity curve under consideration. For each time–activity curve, standard NLR routines (24) were used to derive estimates of VB and rate constants K1, k2, k3, k4. Volumes of distribution (VT) were then calculated with Equation 5:Eq. 5Because NLR is sensitive to noise, for some time–activity curves the method was unable to produce reliable fits, particularly for small VOIs. VT for these VOIs (18/96) were therefore excluded from further analysis.
LGA
In LGA (20), the governing differential equations for the reversible model are rearranged to yieldEq. 6withandFor reversible kinetics, Equation 6 becomes linear after a certain time, when β becomes constant. The slope of the “Logan plot,” α, then represents VT, albeit without correction for the blood volume fraction. Parametric VT images were generated from the dynamic PET data by linear regression fitting to the data in Equation 6, estimating α, for each voxel.
SA
SA (18) assumes that the solution to the governing equations can be described by a weighted sum of a set of predefined basis functions:Eq. 7where the last term is added to correct for blood volume in VOIs. VT and K1 can then be derived from the amplitudes and exponents of the basis functions:Eq. 8andEq. 9Parametric VT, K1, and VB images were generated by fitting Equation 7 to the dynamic PET data, estimating αi and VB values, for each voxel.
Simplified Methods
Of the simplified parameters, we chose to compare the results obtained from the parametric methods with tumor-to-blood ratios (SUVr), as we previously found SUVr to show the highest correspondence with NLR-derived VT (17). SUVr images were derived from the average PET frames acquired at 50–60 min after injection; voxel-by-voxel activity concentrations were normalized to the blood activity concentrations obtained from a VOI within the ascending aorta.
Derivation of Plasma Input Functions
For each patient, a VOI (2–3 mL) within the ascending aorta was defined manually from an early PET frame most clearly displaying the blood pool. The metabolite-corrected image-derived plasma input function (IDIF) was then derived from the corresponding time–activity curve, corrected for both plasma–to–whole-blood ratios and parent fractions obtained from the manually collected arterial blood samples, and calibrated to the whole-blood activity concentrations of the samples collected 50 min after injection as previously described by Verwer et al. (17).
Input Parameter Setting Selection and Validation
For the identification of appropriate input parameter settings, for each patient, VOIs were defined within the largest tumor structure visible on the low-dose CT image (whole tumor; whole tumor excluding edge voxels) and on the averaged image from the last 30 min of 18F-FAZA data (whole tumor; within an area of relatively low activity concentrations; within an area of relatively high activity concentrations). All VOIs were modified manually to exclude large blood vessels visible on early PET frames. Time–activity curves derived from these VOIs were fitted with NLR to obtain VT estimates. Parametric images were generated with LGA and SA for various input settings (Supplemental Tables 1 and 2; supplemental materials are available at http://jnm.snmjournals.org). Averaged VT were obtained by projecting the VOIs onto the parametric images. The input parameter settings yielding the highest correlations between parametric method–derived and NLR-derived VT were then selected for further analysis.
In addition, for assessment of whether the parametric methods were able to accurately quantify the heterogeneous distribution of tracer-specific uptake in tumor tissue, several (4–7) smaller VOIs were defined within areas showing marked heterogeneity. In particular, areas in which voxel-by-voxel differences were observed in VT derived from LGA and VT derived from SA were studied. Input parameter settings were then further refined on the basis of correlations and correspondence of the resulting VT with NLR-derived VT for these VOIs.
After the selection of appropriate input parameter settings for the parametric methods, VOI-based VT, VB, K1, and SUVr estimates were derived for all tumor VOIs with the aforementioned methods and compared with the results of NLR, which was used as the gold standard.
Statistics
Statistical analysis was performed with the SPSS Statistics 20 (IBM) software package. For comparison of the results of the various methods, linear regression analysis was performed. Correlations were assessed with the Pearson correlation coefficient (R2). Correspondence was assessed with the intraclass correlation coefficient (ICC). P values of less than 0.05 were considered significant. Bland–Altman plots were constructed to visualize correspondence and variability. Plots presented in this article were generated with Excel 2007 (Microsoft Corp.).
RESULTS
Patients
Nine patients (3 men and 6 women) ranging in age from 36 to 70 y were included in the study. The median patient weight was 78 kg (range, 59–91 kg), and the median height was 166 cm (range, 158–183 cm). The median size of the largest tumor structure in each patient was 67 mL (range, 9–312 mL). Figure 1 shows 18F-FAZA PET/CT images from a typical patient with non–small cell lung cancer. A typical example of acquired blood data and the resulting IDIF is shown in Figure 2.
Kinetic Analysis and Validation
Parametric images were obtained from LGA and SA at various settings. Figure 3 shows representative VT and SUVr images at the tumor level. Figure 4 shows scatterplots and regression lines for the results of regional comparisons of the various methods. For correlation of these results, 1 VOI was excluded, as the gold standard VT (VT derived from NLR) for this VOI was considered to be unrealistically high (2.06) compared with those for all other VOIs. However, because no obvious reason for the inaccuracy of this data point could be found in PET or blood data, the data point is shown in the scatterplots. For the sake of completeness, scatterplots and regression analysis results for data including this outlier are given in Supplemental Figure 1.
The best results for LGA were obtained with a start time of 30 min (R2, 0.88; ICC, 0.93) (Fig. 4A; Supplemental Table 1). For SA, a set of 30 basis functions with exponents ranging from 0.0175 to 1.9 min−1 yielded the best results (R2, 0.79; ICC, 0.81) (Fig. 4C; Supplemental Table 2). Better correlations were achieved by selecting different boundaries for the basis function exponents (e.g., an R2 of 0.87 for exponents ranging from 0.01 to 3.0 min−1). However, this approach also changed the slope of the regression line and decreased the absolute correspondence with NLR-derived VT (ICC, 0.76). VB values obtained from SA showed significant correlations with corresponding NLR results (R2, 0.89; ICC, 0.93). K1 values showed no correlation (R2, 0.09).
Figure 4B shows comparisons of SUVr with VT (R2, 0.85; ICC, 0.80). Figure 4D shows the relative differences between VT derived from LGA and VT derived from SA as a function of VB. A typical example of the relative differences at the voxel level is shown in Figure 3E. Figure 3F shows the relative differences between SUVr and SA-derived VT. Bland–Altman plots for the various methods compared with NLR are given in Supplemental Figure 2.
DISCUSSION
18F-FAZA is a nitroimidazole-based PET tracer for noninvasively assessing tumor tissue pO2 levels. For 18F-FAZA to be used for this purpose, accurate quantification of specific tracer uptake is required. The present study validated the use of 2 parametric methods for quantification at the voxel level.
All methods were compared with NLR. Because NLR is sensitive to noise, the comparison was VOI based. This approach could lead to quantification errors in areas with marked variations in PET signal because such variations could be averaged out. Because hypoxia is expected to be heterogeneously distributed within a tumor, so is 18F-FAZA uptake. To take possible errors due to heterogeneity into consideration, we included in the analysis multiple smaller VOIs that were defined within areas of the tumor showing various characteristics (such as relatively high or low PET signal or differences in VT derived from the 2 parametric methods). The results for these smaller VOIs were found to be consistent with those for the other VOIs.
Both parametric methods produced VT showing high correlations with those obtained from a full kinetic analysis. However, as shown in Figures 3 and 4, some differences between the methods were observed. Figures 3D and 3E indicated that these differences might be explained by high blood volume fractions within voxels because LGA does not incorporate blood volume correction. Regional analysis, however, showed no correlation of VB with differences between VT derived from LGA and VT derived from SA.
LGA performed well (R2, 0.88; ICC, 0.93), although some increased variability was observed at relatively low as well as relatively high VT. A possible explanation is that even though a start time of 30 min generally led to the highest correlation with NLR-derived VT, for some VOIs, particularly those with high VT, the second term in Equation 6 might not yet have been constant at that start time. The observed bias at high VT can indeed be decreased by selection of a later start time (e.g., 40 min after injection) (Supplemental Fig. 3), but this approach leads to errors for other VOIs. Another explanation is noise-induced bias. LGA is known to have noise-induced bias, particularly at high VT, causing underestimation (25). LGA performance could be somewhat improved (R2, 0.90; ICC, 0.94) by fitting Equation 6 with orthogonal regression analysis (26).
SA results showed high correspondence with NLR-derived VT (R2, 0.79; ICC, 0.81). The disadvantages of SA are the sensitivity of the method to minor changes in input parameters and long computing times (∼10 times longer than LGA). As shown in Supplemental Figure 2, SA results were more variable than LGA results; this finding was expected because more parameters need to be estimated in SA. However, SA also has advantages. VB values correlated highly with those obtained from NLR (R2, 0.89). Calculating VB may be useful as a quality assurance step and for providing information on tumor vascularity. Estimates of K1 by SA and NLR were markedly sensitive to noise. For the final selection of input parameters, no correlation was found between SA-derived and NLR-derived K1 values (R2, 0.09) when all VOIs were considered. Correlation was improved when the analysis was limited to larger VOIs (R2 for VOI sizes of >10 mL, 0.52), illustrating the influence of noise on estimates of K1 with SA.
Both parametric methods required an elaborate imaging protocol consisting of dynamic PET imaging and blood sampling. For 18F-FAZA, use of the IDIF in combination with venous blood sampling instead of (continuous) arterial blood sampling has been validated (17), simplifying the protocol. However, the need for venous blood sampling and metabolite analysis may still limit the clinical applicability of the presented methods. In the study of Verwer et al. (17), metabolite formation was observed to be slow, plasma-to-blood ratios were typically equal to 1, and both measures showed limited interpatient variability. Indeed, omission of the corrections for plasma-to-blood ratios and parent fractions led to an underestimation of VT by only 8%. These findings need to be further substantiated under various clinical conditions but hint at the possibility of using the whole-blood IDIF, which would improve the clinical feasibility of the acquisition protocol.
Nevertheless, in routine clinical settings, a static imaging protocol may be preferable because dynamic imaging generally requires longer imaging times and may therefore be more costly and more prone to artifacts related to patient motion. Therefore, SUVr was also investigated. For calculation of SUVr, typically a short-duration (e.g., 10-min) static PET/CT scan is acquired at a late time point (>50 min after injection). Tumor tissue and blood activity concentrations are then derived from the PET image directly. The blood pool will not be clearly visible on the late PET image, but this characteristic will not be a limitation with respect to deriving blood activity concentrations because the ascending aorta and aortic arch are clearly identifiable on the low-dose CT image. However, it can lead to considerable quantification errors caused by the blood volume in the tumor tissue VOIs because SUVr does not correct for VB (Fig. 3F). In the present analysis, VOIs were manually adjusted to exclude areas of markedly high VB (i.e., visible blood vessels) by use of early PET frame data, thereby reducing potential errors caused by high VB. With a static imaging protocol, this quality assurance step cannot be performed. In future applications, this limitation of SUVr should be considered. In the present analysis, regional SUVr estimates showed high correlations with NLR-derived VT (R2, 0.85) but significant negative bias, particularly at high VT.
Which method should be used in clinical settings ultimately depends on the clinical objective. For assessing whether tumor tissue is hypoxic or not, defining a suitable threshold for the quantification parameter used could be sufficient. Depending on the threshold, systematic bias and underestimation at high VT (like those observed with SUVr) might not affect clinical conclusions. In that situation, SUVr would be the method of choice because both the method and the acquisition protocol are straightforward. Large clinical trials relating quantification parameter values to clinical outcomes are needed to assess whether the observed bias in SUVr results in clinically relevant quantification errors. For purposes requiring more detailed information, such as radiation dose painting and early response monitoring (for which the ability to assess minor variations is essential), accurate voxel-by-voxel quantification is required. In such scenarios, parametric methods are recommended. SA is more versatile than LGA but shows larger variability. On the other hand, LGA results may be affected by the lack of blood volume correction. However, the effect will be less pronounced for LGA than for SUVr because with LGA, early PET frame data are available for excluding large blood volume regions.
Even though the parametric methods performed well in the present study, clinical conditions are likely to change over the course of a response monitoring study. Longitudinal trials are needed to evaluate the robustness of the methods under various clinical conditions. It would also be of interest to directly compare 18F-FAZA voxel VT derived from the parametric methods with localized pO2 needle electrode measurements to assess the relationship between VT and tissue pO2 levels. However, such a study would be invasive and feasible only for superficial tumor structures.
CONCLUSION
Both LGA and SA produced accurate VT images for the quantification of 18F-FAZA. Additionally, the latter also produced accurate VB values. SUVr also correlated well with NLR-derived VT but showed significant bias, particularly at high VT.
DISCLOSURE
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. This study was performed within the framework of the Center for Translational Molecular Medicine (AIRFORCE project; grant 03O-103). No other potential conflict of interest relevant to this article was reported.
Acknowledgments
We thank our colleagues at the Department of Radiology and Nuclear Medicine and all patients who participated in the study.
Footnotes
Published online Oct. 16, 2014.
- © 2014 by the Society of Nuclear Medicine and Molecular Imaging, Inc.
REFERENCES
- Received for publication April 17, 2014.
- Accepted for publication August 11, 2014.