Estimation of Tumor Volumes by ¹¹C-MeAIB and ¹⁸F-FDG PET in an Orthotopic Glioblastoma Rat Model

DISCUSSION

In this study, histology- and PET-derived tumor volumes of 2 phenotypically distinct and different brain tumor types in an orthotopic GBM rat model were compared. Two different radiotracers—the widely used ¹⁸F-FDG reflecting the metabolic rate and the amino-acid tracer ¹¹C-MeAIB reflecting the rate of the system A amino-acid transport—were tested. PET images were reconstructed with OSEM2D and MAP3D algorithms, and a parametric image of the ¹¹C-MeAIB MAP3D-reconstructed images was calculated. Image segmentation was performed with fixed thresholding of VImax and MBI. Our aim was to identify the best combination to determine tumor volume by PET.

As reference, histology-derived tumor volumes were used, although doing so presents some inherent limitations. On formaldehyde fixation, the tissue is known to shrink approximately 10%. Assuming that this was the case for all the tumor specimens, the shrinkage ought not to have affected histologic reference volumes after PET scanning, because these presumably were smaller than the tumors were in situ. Moreover, histologic tumor volume was determined manually and was therefore potentially prone to variability. However, when tested, this variability was very small, with narrow 95% LOAs and high R² values.

All scans were performed as dynamic scans to extract time–activity data. From this, the optimal 20-min scan interval, defined as the interval with the highest T/B ratio, was identified. Previous data on ¹⁸F-FDG have shown that extending the time from injection of tracer to acquisition of scan—so-called delayed imaging—improves the T/B ratio (2). Our data support this, in particular for ¹¹C-MeAIB (Figs. 2 and 4). However, it is important to note that the SEM of the T/B ratio increased also (Fig. 4) because of the declining number of annihilations taking place due to radioactive decay. Dynamic time–activity data for ¹¹C-MeAIB have not been previously reported. Interestingly, one study based on autoradiography and using the related, but less specific, tracer α-aminoisobutyric acid in rats with induced subcortical rat gliomas reported T/B ratios of 1.8–6.8 (12). In our study, T/B ratio in the U87MG tumors was 7.6 and in the T87 tumors was 3.8. The reason for this twofold difference is probably the pronounced difference in tumor composition. The U87MG tumors are very dense and consist almost exclusively of tumor cells in contrast to the highly infiltrating T87 tumors that, to a certain extent, like the GBMs seen in humans, are composed of both tumor cells and brain parenchyma. The latter cell composition obviously decreases the T/B ratio, because ¹¹C-MeAIB is hardly taken up by brain parenchyma (Fig. 2B). In summary, our time–activity data support the use of delayed imaging in that it increased the contrast between tumor and the surrounding brain parenchyma. Because no real plateau phase was seen in the ¹¹C-MeAIB time–activity data (Fig. 4), we cannot exclude that further delay in imaging would increase the T/B ratio even more, but at the expense of an increase in statistical noise.

From the ¹¹C-MeAIB time–activity data, a notable difference in tumor and background voxel intensities was observed. After the initial vascular blood flow phase, the tumor tissue showed an almost linearly increasing tracer uptake, whereas normal brain tissue (background) exhibited a minimal uptake throughout (Fig. 2B). Therefore, a parametric image based on the last five 10-min frames was calculated. The derived parametric image represents the kinetic behavior of each single voxel in just one image; in principle, it has the same properties as dual-time-point imaging (13) but with a better statistical behavior because more frames are used in the calculations. In the noninfiltrating U87MG tumors, the agreement between histologically and parametrically derived tumor volumes was inferior to that of standard static PET images (Table 1). In the infiltrating T87 tumors, parametric and standard static PET imaging performed almost equally well (Table 2); but when observer agreement was taken into account (Fig. 7), parametric imaging was optimal for T87 tumors. The best threshold values for the parametric images were remarkably low (0.1% of SUV_max and 1.0 × MBI, Table 2). The latter indicates that some areas of the tumors did not take up enough ¹¹C-MeAIB over time to produce a significant increase in voxel intensity and therefore did not appear on the parametric images. These tumor areas seemed to be the peripheral parts of the tumors when we compared parametric and standard static images. In addition, the parametric imaging technique required long dynamic scans (1 h), and the postprocessing was time-consuming. These factors make this approach less feasible for routine use in a preclinical—and particularly in a clinical—setting, and calls for more automated processing.

We tested 2 different reconstruction algorithms because previous data indicated that MAP3D reconstruction yields superior spatial resolution in the entire field of view compared with OSEM2D (14). Because we scanned 2 animals simultaneously, their heads were not located in the center of the field of view; therefore, MAP3D should a priori be superior in our specific setup. However, as seen from the results in Tables 1 and 2, MAP3D was not consistently superior to OSEM2D reconstruction.

To segment the PET images we applied thresholding, which is just one of the at least 4 broad image segmentation techniques: thresholding (15), variational (16), learning method (17), and stochastic modeling (18). All of these segmentation techniques have pros and cons (6). We chose the thresholding technique because, along with visual delineation performed by experts (19), it is the most frequently used technique (6). Fixed thresholding of VImax and MBI was applied. Two confounding factors using these techniques are variations in tumor size (20) and tumor inhomogeneity (21), both of which may influence the optimal thresholding value (22). Whether any of the more advanced, but less frequently used, segmentation techniques would have improved our results remains unanswered; this issue would be of interest to address in a future study.

From the histologic tumor volume data, the U87MG tumors were clearly smaller than the T87 tumors. The reason for this was that rats were euthanized when tumors were depicted on PET, which happened earlier with the noninfiltrating U87MG tumors than with the infiltrating T87 tumors. On the basis of the small size of some of the U87MG tumors, we would expect partial-volume effects to affect their PET measurements. Because partial-volume effects will make small tumors appear larger (23), we would expect a volume overestimation of the smaller tumors in the Bland–Altman plots. In fact, this was the case in tumors smaller than 40 mm³, because 5 of 7 with ¹⁸F-FDG and 7 of 7 with ¹¹C-MeAIB were indeed estimated to be larger than their histologic counterparts (Fig. 5). Still, the overall volume estimation of the U87MG tumors (Figs. 5B and 5C) and the observer agreement (Fig. 6) were good with narrow LOAs. The best combination was ¹¹C-MeAIB imaging reconstructed by MAP3D and image segmentation by fixed thresholding of VImax. Surprisingly, the intraobserver agreement was a little lower than the interobserver agreement (Fig. 6), albeit without an obvious explanation.

Delineating the T87 tumors proved difficult, as expected, because both the ¹⁸F-FDG and the ¹¹C-MeAIB scans yielded very wide 95% LOAs regardless of reconstruction algorithm and image segmentation (Table 2). As already described, the better combination, when including both volume estimation (Fig. 5F) and observer agreement (Figs. 7C and 7D), was ¹¹C-MeAIB parametric imaging with fixed thresholding of VImax. The LOA for this combination was −44.8 to 58.9 mm³ for tumors with a median volume of 144 mm³—that is, an approximate deviation of plus or minus 30%—which is hardly acceptable. To our knowledge, comparable studies of histology- and PET-derived brain tumor volume do not exist, but studies comparing MR imaging– and histology-determined brain tumor volumes do. In one study with a mouse brain tumor model, the correlation between MR imaging– and histology-derived tumor volumes showed an R² value of 0.85 (24). In another, a syngeneic rat glioma model was used, and an R² value of 0.76 was obtained (25). The respective R² values in our study were, for U87MG tumors, 0.98 with ¹⁸F-FDG and 0.99 with ¹¹C-MeAIB, whereas for T87 tumors, the corresponding values were 0.81 and 0.79, respectively. Based on these correlation analyses, PET determination of tumor volumes seems comparable to or even better than MR imaging–based tumor volume assessment. To address this thoroughly, a head-to-head comparison is needed.

Taken together, our T87 results obtained in a relatively small number of animals indicated that ¹¹C-MeAIB was superior to ¹⁸F-FDG when considering both volume and observer agreement. This result can be explained in part by the superior tumor-specific uptake of ¹¹C-MeAIB (4) compared with ¹⁸F-FDG, which is known to be taken up in different areas of the normal brain, including the cortex, the basal ganglia, and the thalamus (26). This advantage underlines the potential value of tumor-specific amino acid tracers such as ¹¹C-MeAIB for the future imaging of brain tumors.

For neurosurgeons and radiation oncologists, it is critical to have borders of GBMs delineated before tumor resection or radiation; our results do not support the use of PET for this purpose. But for the neurooncologist, molecular changes on the cellular level, which normally precede the MR imaging–detected structural changes (27), might be an optimal determinant of early treatment response. Potentially, this ability could minimize the times at which individual patients receive ineffective and often toxic treatment. Future studies using clinically relevant orthotopic models such as the model used in our study, as well as clinical studies in patients, need to be performed to address this aspect.