Methods to Monitor Response to Chemotherapy in Non-Small Cell Lung Cancer with ¹⁸F-FDG PET

MATERIALS AND METHODS

Scans were performed on 2 separate groups of patients. First, a group of 10 patients (8 men, 2 women; mean age, 53 ± 6.8 y) with NSCLC stage IIIB/IV was scanned twice on consecutive days before the start of chemotherapy to assess test–retest variability. Second, in a separate group, 30 randomly selected dynamic scans were used, which were obtained on 19 patients (15 men, 4 women; mean age, 59.4 ± 8.1 y) with NSCLC stage IIIA-N2, as part of an ongoing response-monitoring study.

Scans were performed using a state-of-the-art PET scanner (ECAT EXACT HR+; Siemens/CTI, Knoxville, TN). This scanner has an axial field of view of 15 cm, divided into 63 contiguous planes. The patient was positioned supine on the scanner bed with the tumor in the center of the axial field of view.

All patients fasted for 6 h before scanning. Patients received 2 venous catheters: one for injection of ¹⁸F-FDG contralateral to the tumor and the other for venous blood sampling. Acquisition started with a 10- to 15-min transmission scan to correct for photon attenuation (9), followed by a bolus injection of 370 MBq ¹⁸F-FDG in 5 mL saline through an injector (Medrad International, Maastricht, The Netherlands) at 0.8 mL/s, after which the line was flushed with 42 mL saline (2.0 mL/s). Simultaneous with the injection of ¹⁸F-FDG, a dynamic emission scan (in 2-dimensional mode) was started with a total duration of 60 min with variable frame length (6 × 5 s, 6 × 10 s, 3 × 20 s, 5 × 30 s, 5 × 60 s, 8 × 150 s, and 6 × 300 s). All dynamic scan data were corrected for dead time, decay, scatter, randoms, and photon attenuation and were reconstructed as 128 × 128 matrices using filtered backprojection (FBP) with a Hanning filter (cutoff, 0.5 cycle/pixel). This resulted in a transaxial spatial resolution of around 7-mm full width at half maximum. Before injection of ¹⁸F-FDG, a blood sample was collected for determination of the plasma glucose level. In addition, 3 venous blood samples were drawn at 35, 45, and 55 min after ¹⁸F-FDG injection as quality control for the image-derived input function (10,11) and for plasma glucose measurement (hexokinase method, Hitachi 747; Boehringer Mannheim, Mannheim, Germany).

RESULTS

The 3k model provided significantly better fits than the 4k model in 26 (87%) and 27 (90%) of 30 scans according to Akaike (19) and Schwarz (20) criteria, respectively. In other words, the data did not support inclusion of a fourth rate constant and, consequently, it was set to zero in this study.

The mean value ± SD for the plasma glucose level found in this study was 5.3 ± 0.6 mmol/L (range, 3.9–6.5 mmol/L).

NLR using the standard (3k) 2-tissue compartment model with a blood volume component and using an image-derived input function proved to have an excellent test–retest variability (ICC, 0.95; 95% confidence interval [CI], 0.82–0.99). The inter- and intraobserver variabilities were also very good (ICC, 0.98; 95% CI, 0.96–0.99) for both. It was concluded that NLR using the above-mentioned model, with an image-derived input function, could indeed be used as a gold standard. The mean value ± SD for glucose consumption using NLR with 3 rate constants (3k model) was 0.15 ± 0.08 μmol/mL/min (range, 0.02–0.33 μmol/mL/min).

Of the total of 34 alternative methods investigated, 24 had a suboptimal correlation with NLR (r² < 0.95; Table 1). The best correlations with NLR were found with the SUV corrected for body surface area (BSA) and plasma glucose (SUV_BSAg) measured at 40–60 and 50–60 min, the SKM at 40–60 and 50–60 min, the net influx constant method corrected for BSA, and the Patlak graphical analysis (Table 1; Fig. 1). The best scanning period for Patlak graphical analysis was found to be from 10 to 60 min.

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 1.

Scatter plots of correlation between NLR and SUV_BSAg (A), simplified kinetic method (B), and Patlak graphical analysis (C).

In Table 1 the slope and intercept of the regression lines are given, allowing NLR results to be converted to any of the other methods and vice versa.

DISCUSSION

¹⁸F-FDG PET appears to be a promising technique to monitor response to chemotherapy (8). Unfortunately, its true value is unknown, as no meta-analysis of reported data is feasible because of the variety of analytic methods used in different studies. Various approaches have been used, ranging from visual assessment (qualitative), through semiquantitative indices, to full kinetic analyses of ¹⁸F-FDG uptake (7). It is by no means clear whether these methods have similar sensitivity for monitoring changes after therapy. Because interest in response monitoring is growing, selection of the most appropriate method is vital to determine the role of ¹⁸F-FDG PET in this field.

For each tumor the presence and possible effect of a fourth rate constant (k₄) should be assessed to make a proper decision on whether to include a k₄ parameter in the model (22,23). In this study on NSCLC, a model with 3 rate constants provided significantly better fits than other models that included a fourth rate constant.

Various analytic methods were compared in our study by applying them to the same dataset. One question was whether NLR, generally accepted as the gold standard method, could still be considered as such when an image-derived input function is used. For full kinetic modeling (NLR), both dynamic scanning (tissue time–activity curve) and arterial sampling (arterial plasma time–activity curve) should be used. Arterial cannulation is less suitable for routine clinical response studies, where repetitive scans are required. Recently, measurement of the metabolic rate of glucose (MR_glu) using an image-derived input function obtained from thoracic vascular structures has been validated against arterial blood sampling (11). In this study an image-derived input function was used, applying quality control measures as defined previously (10). However, manual ROI definition of the vascular structures (aorta, atrium, or ventricle) could introduce inter- and intraobserver variability. In addition, noise could be introduced by the limited number of counts acquired in each frame. Nevertheless, the ICC for both test–retest and intra- and interobserver variability was excellent, indicating that even when an image-derived input function is used NLR is a reproducible method for measuring glucose consumption. Therefore, it was used in our study as the gold standard for assessing other simplified methods.

Results of the other (simplified) analytic methods were correlated with NLR (Table 1; Fig. 1) and a cutoff value of r² > 0.95 for selected methods that approach NLR. Although test–retest and inter- and intraobserver variation could also have been determined for the simplified methods, this analysis did not seem to be very useful. The main variability in NLR will be caused by variation in defining ROI over the blood pool for the image-derived input curve. By combining as many ROIs as possible over 3 different blood-pool structures, these variations can be minimalized (10). For simplified methods such as SUV, only tumor ROIs need to be defined and, as a semiautomatic program is used, variation will be much less than for NLR. Test–retest data only address variability for unchanged conditions. However, uptake of ¹⁸F-FDG in the tumor and distribution of ¹⁸F-FDG throughout the body may be expected to change because of the effects of therapy. These changes are taken care of in NLR (for each study a new input function is defined) but are not necessarily incorporated in simplified methods (24,25). Therefore, quantitative accuracy (i.e., correspondence with glucose consumption [NLR]) appears to be more relevant for response-monitoring studies.

As expected in the case of no dephosphorylation (k₄ = 0), Patlak graphical analysis approaches the accuracy of NLR with the highest correlation obtained for the interval from 10 to 60 min after injection. Although this should be investigated for each tumor type, it is clear that, for NSCLC, Patlak graphical analysis is an excellent alternative for NLR. This would allow for a simpler scanning protocol (less frames) and the option to generate functional glucose consumption (MR_glu) images.

For routine clinical use, an even simpler and shorter scanning protocol (single frame) would be preferable. However, methods using such a protocol have been criticized because they are based on several simplifying assumptions, which may not be valid in repeated (response monitoring) studies and thus may introduce errors in the final analysis (7). As for the semiquantitative SUV, correcting for BSA and plasma glucose appeared to be preferable over correcting for body weight or lean body mass in agreement with previous studies (25–27). Our findings contradict the recommendations made by the EORTC PET study group (8). They did not recommend correction for plasma glucose because of concern about the accuracy of measurements in many institutes. Our study indicates that an accurate correction should be used.

The net influx constant method described by Sadato et al. (18) also showed a good correlation with NLR. However, this method is directly proportional to SUV because it uses a fixed scaling factor for translating SUV into a measure of MR_glu. The simplified kinetic method of Hunter et al. (17) showed even better agreement with NLR than with SUV. The method holds promise for routine clinical applications because it is requires only a static scan and a venous blood sample for the tail of the plasma input function (without having to measure the plasma curve itself).

The analytic methods used in ¹⁸F-FDG PET studies are simplifications of the underlying physiology. Variability in results can be induced by many factors, most of which can potentially be avoided at the cost of more complicated study protocols. Simplifications in the models can be implemented, however, at the cost of accuracy. This is important to keep in mind when selecting an analytic method for a study protocol. Selection of the degree of accuracy needed (i.e., method required) will depend on how large the differences are between responding and nonresponding tumors.

It is not possible to select the optimal method for response monitoring in NSCLC using the results of our study. This will be possible only in a larger series of patients where response data are compared with clinical outcome because selection of the optimal trade-off between accuracy and simplicity will depend on the actual changes being measured. The purpose of our study was primarily to reduce the multitude of available methods to a limited number of potentially worthwhile techniques, each with a different degree of complexity regarding correction factors used and scanning protocols. We suggest that the following methods be compared with NLR in a prospective clinical study: Patlak graphical analysis from 10 to 60 min, the simplified kinetic method at 40–60 min, and SUV corrected for BSA and plasma glucose at 40–60 min. Such a study is currently in progress in our institutions.

CONCLUSION

Even with an image-derived input function, NLR has excellent test–retest and observer agreement supporting its use as a gold standard method. Of a total of 34 potential analytic methods, 10 showed good correlation with NLR (r² > 0.95). By taking into account the degree of complexity of the methods, only 4 remain. The actual value of these methods needs to be evaluated in a large prospective response-monitoring study that compares results with patient outcome.

For response-monitoring studies in NSCLC, the nomograms in this study also provide the possibility of converting results obtained by one analytic method to another. This will allow comparison across studies (i.e., meta-analysis).