Quantification of Hypoxia in Human Glioblastoma using PET with 18F-FMISO

Abstract

Purpose

This study aimed to investigate the results of compartmental modeling (CM) and spectral analysis (SA) generated with dynamic 18F-FMISO tumor images. Besides, the regular tissue-to-blood ratio (TBR) images were derived and compared with the dynamic models.

Methods

Nine subjects with glioblastoma underwent PET/CT imaging with the 18F-FMISO tracer. The protocol for PET imaging began with 15 min in dynamic mode and two 10-min duration static images at 120 min and 180 min post-injection. We used the two-tissue compartmental model for CM at the voxel basis, and we conducted SA to estimate the 18F-FMISO accumulation within each voxel. We also investigated the usual tumor-to-blood ratio (TBR) for comparison.

Results

The images of the tumor showed different patterns of hypoxia and necrosis as a function of PET scanning times, while CM and SA methods based on dynamic PET imaging equally located tumor hypoxia. The mean correlation of K_i images of all subjects between CM and SA was 0.63 ± 0.19 (0.24-0.86). CM produced less noisy K_i images than SA, and, in the contrary, SA produced accumulation component images more clear than with CM. CM-K_i and SA-K_i images were correlated with TBR images (r = 0.72 ± 0.20 and 0.56 ± 0.26, respectively). In the only subject having a continuously increasing tumor time-activity curve, the k₃ image showed a high uptake in the necrosis region which was not apparent in TBR or K_i images.

Conclusion

Based on these results, the combination of CM and SA approaches was found more appropriate in generating voxel-based hypoxia images.

Introduction

Parametric images in positron emission tomography (PET) provide quantitative maps of drug pharmacokinetic parameters from dynamic images. Various physiological parameters can be extracted from parametric images such as blood flow, glucose metabolism, enzyme activity, protein concentrations, and hypoxia [1].

The standard acquisition technique for parametric images uses the dynamic time series of a tracer. The information is generated by analyzing the time-activity curve (TAC) of the region-of-interest (ROI) or the TAC of individual voxels in the image. The fitting at the voxel level is generally noisy, and the process is time-consuming. However, many techniques have been developed to overcome the noise, such as applying weight factors, optimizing the models using noise reduction techniques, and implementing the kinetic data inside the image reconstruction process [1].

Several methods have been used to generate parametric images from different radiotracers. The most notable method is the compartmental kinetic modeling (CM) [2]. CM estimates the physiological kinetic parameters based on a model construction of inputs and outputs from which it retrieves the rate constants. Various quantitative parameters can be estimated from CM, such as the net influx rate K_i, the perfusion K₁, and tissue blood volume (TBV). In addition, similar methods based on CM estimate such parameters, for instance, the graphical analysis techniques such as Patlak or Logan analysis estimate K_i or glucose metabolism [3, 4]. Another method used for generating parametric images is spectral analysis (SA) [5, 6]. SA is a technique used to decompose the PET dynamic TACs into their principal compartments or components. The outputs of this analysis are correlated to physiological parameters. The most potent advantage of SA over CM is that SA is used without a priori information or assumptions on the model.

Tumor hypoxia is one of the characteristics of most cancers. It is defined as a deficiency in blood supply and thus of oxygen inside the tumor. In terms of imaging, the tumor becomes less perfused. A variety of PET tracers have been used to detect tumor hypoxia mainly 18F-FMISO and 18F-FAZA [7]. Evaluation of PET images for glioblastoma multiforme (GBM) hypoxia with 18F-FMISO was mainly based on the type of acquisition. Static PET images were primarily using the standard uptake value (SUV) and the ratio method [8]. These techniques did not give the complete kinetic information about hypoxia's physiology in GBM, as they have essentially images acquired at a specific time. Therefore, clinicians nowadays prefer the use of dynamic images, which can give more information about hypoxia. In 1995, Casciari et al. designed a model for 18F-FMISO dynamic images [9]. The main aim was to estimate the 18F-FMISO reaction rate constant KA. This parameter was related to oxygenation inside cells. The model was able to estimate the parameter, but the accuracy was affected by other intrinsic parameters with the model. In another study, Bruehlmeier et al. used several models for 18F-FMISO with GBM [10]. They found distribution volume was the best parameter at early 18F-FMISO images, but no correlation at late images. Finally, Thorwarth et al. applied a new model based on weighting factors, representing the perfusion, accumulation, and blood inside tumors [11, 12]. The estimated parameters were superior to SUV but were not able to predict the treatment outcome. However, further investigation of these techniques is important, as defining tumor hypoxia increases diagnostic accuracy, consequently enabling cancer treatment personalization.

This study’s objective was to investigate the results of compartmental modeling (CM) and spectral analysis (SA) generated with dynamic 18F-FMISO images from glioblastoma patients. In addition, a comparison of TBR images with CM and SA was also studied.

Materials and Methods

Imaging Data

The project was composed of 9 subjects diagnosed with grade IV GBM. We applied the Lim and Berridge protocol for 18F-FMISO preparation [13]. Dynamic PET scan 18F-FMISO was utilized for 15 min (activity 3.7 MBq/kg). The used time frames were 12 × 10 s, 8 × 15 s, and 11 × 60 s. Besides, two static scans were done at 2 h and 3 h post-injection. PET images were reconstructed using standard techniques (3D-OSEM ordered-subset expectation-maximization algorithm.). All PET scans were conducted on Discovery RX (General Electric Medical System). Computed tomography (CT) scan (120 kVp, 80–250 mA) was used in the beginning for attenuation correction, anatomical localization, and co-registration between dynamic and static scans. Image analysis on the corrected PET images began with calculating TBR images at 3h with a 1.4 threshold level. Next, we did a kinetic modeling and spectral analysis on each voxel for all subjects. We used r-squared to compare the results of fitting curves of CM and SA. Parametric images of K1, k2, k3, and ki and volume of distribution (VD) were estimated using CM and SA for comparisons. The project was performed according to the ethical standards of the University of Bordeaux-II, and informed consent was collected from all subjects.

Compartmental Modeling

The blood input function (IF) was derived from the PET images of the common carotid artery by drawing an ROI on the sagittal view to encompass a large area of the carotid, which was less prone to the partial volume effect (PVE). In this study, we assumed the plasma input function was identical to the whole blood input function, and no metabolites existed in the plasma. Spillover and PVE were minimized by placing the carotid artery ROI at least 2 voxels away from the high uptake regions. The IF was used for both CM and SA.

The mathematical model for CM was the two-tissue compartment model as shown in Fig. 1 [2]:

$${\mathrm{TAC}}_{\mathrm{PET}}\left(t\right)=\frac{K_1}{{\alpha }_2-{\alpha }_1}\left[\left(k_3+k_4-{\alpha }_1\right)e^{-{\alpha }_{1}t}+\left({\alpha }_2-k_3-k_4\right)e^{-{\alpha }_{2}t}\right]\otimes \text{IF}\left(t\right)+\mathrm{TBV}\times \text{IF}\left(t\right)$$ 1

where

$${\alpha }_1=\left[k_2+k_3+k_4-\sqrt{{\left(k_2+k_3+k_4\right)}^2-4k_2{k}_4}\right]/2$$

$${\alpha }_2=\left[k_2+k_3+k_4+\sqrt{{\left(k_2+k_3+k_4\right)}^2-4k_2{k}_4}\right]/2$$

Fig. 1.

The compartmental model consisting of three compartments. C_B is the tissue blood volume in the ROI or voxel, C_F is the free radiotracer compartment, and C_A is the compartment of the accumulated bound radiotracer

K₁, k₂, k₃, and k₄ are the rate constants. In this study, we assumed k₄ = 0 and IF derived from the carotid artery images as whole blood. The net uptake rate K_i was estimated as K₁k₃/(k₂ + k₃).

Basically, (1) is composed of three terms representing the three compartments shown in Fig. 1. The accumulation compartment of CM to be compared to the accumulation component obtained with SA is expressed as, by considering k₄ = 0, and by considering the input response function without the convolution [2]:

$${\mathrm{C}}_{\mathrm{A}}\left(t\right)=\frac{K_1{k}_3}{k_2+k_3}(1-exp\left(-\left(k_2+k_3\right)t\right)$$ 2

Spectral Analysis

We applied the SA technique to analyze the PET TACs. It allows us to decompose the TACs into their expected components. These components were assumed to be blood, perfusion, and accumulation. Therefore, this decomposition can help us isolate the accumulation component, which can be related to the hypoxic tumor. In contrast to compartmental modeling, the main advantage of SA is that it does not need prior assumptions about the number of components. SA can determine the amplitudes A_i and the exponents B_i that can give the best fit to the TACs curve. The main equation of SA was given as [14]:

$$\mathrm{TAC}\left(t\right)=\sum _{i=1}^N{A}_i{e}^{-B_{i}t}\otimes \text{IF}\left(t\right)+A_{N+1}\text{IF}\left(t\right)$$ 3

where A and B in (3) were assumed to have real non-negative values. The B values ranged from 10⁻⁵ to 1 (we equivalently refer to B with Log10(B) and thus, B varies from - 5 to 0). The number of exponentials was set to N = 100. However, the maximum number of SA components with our subjects was 6. The second term in (3), A_N + 1IF(t), accounts for tissue blood volume equivalent to TBV × IF(t) in (1). The outputs of SA were reported to correspond to the physiological parameters [6, 14–16]. These parameters were the perfusion K₁, the net uptake of the tracer K_i, and the volume of distribution (VD). In this study, we calculated the 18F-FMISO K_i = K₁k₃/(k₂ + k₃) as equal to the amplitude A_i at B = 0 [15].

To compare the accumulation component of SA to that of CM as in (2), the SA B values encompassing the values from B₁ =  − 5 to a value B_lim were used in the following equation, ignoring the convolution with IF, providing the input response function of the accumulation component:

$$C_A\left(t\right)=\sum _{B_1}^{\begin{array}{c}{B}_{\lim }\\ \end{array}}{A}_i{e}^{-B_{i}t}$$ 4

where B_lim was incrementally determined such that (2) and (4) are the mostly correlated curves.

Tissue-to-Blood Ratio

TBR parametric images were calculated as the ratio of the radiotracer concentration in the tumor to its concentration in blood taken from IF at the specific time [17]. To discriminate hypoxia in the TBR images, we used the conventional 1.4 thresholds on TBR images [17].

Statistical Analysis

Values from CM, SA, and TBR were compared with correlation coefficients and with t test.

Results

Figure 2 shows the distribution of 18F-FMISO in the brain of the same subject during the same PET session at four different times post-injection. The area of the tumor and necrosis differed among the times, indicating the uptake and washout progress of 18F-FMISO inside the tumor. Since the accumulation of 18F-FMISO in the tumor is continuously varying with time, in the clinic, it is usual to consider the latest image even at 4 h post-injection to locate hypoxia which is supposed having less exchange of 18F-FMISO with blood. Greater changes were noticed in the earlier images post-injection, while, based on the whole TAC, the tumors from the 9 subjects were shown to be perfused or highly hypoxic as shown in Fig. 3. In this figure, subject 1 (S1) had an accumulating TAC reflecting a highly hypoxic tumor, subject 2 (S2) had a declining TAC reflecting a perfused tumor, and subject 3 (S3) had a stable TAC. Thus, relying on a single image to locate hypoxia is not a determinant.

Fig. 2.

The uptake of 18F-FMISO in subject 1 (S1) measured at four different times showing the progress of 18F-FMISO accumulation in the tumor

Fig. 3.

The time-activity curve of the tumor from the 9 subjects. Subject 1 (S1) with an accumulating TAC reflecting a highly hypoxic tumor, subject 2 (S2) with a declining TAC reflecting a perfused tumor, and subject 3 (S3) having a stable TAC

Table 1 shows the goodness of fit as calculated using r-squared. The values of K_i,K₁,VD, and TBR from CM and SA were compared in Fig. 4 as parametric images in the nine subjects having different tumor behavior. In all subjects’ images, CM-K_i tumor images appeared well defined and less noisy. The images of K₁ among CM and SA look similar, while the VD images were not correlated. Figure 5 shows the values of K_i, K₁, and VD in the whole tumor from the 9 subjects for CM, SA, and TBR as boxplots where the values were normalized to their respective means to relatively compare the dispersion of the values in each group. The t test between these groups taken two-by-two was not statistically significantly different (p < 0.05). The correlation of K_i value within the whole tumor volume between the CM and SA methods in all subjects was 0.63 ± 0.19 (0.24-0.86). CM-K_i and SA-K_i images were correlated with TBR images (r = 0.72 ± 0.20 and 0.56 ± 0.26, respectively). In addition, the correlation of K₁ value between the CM and SA methods in all subjects was 0.85 ± 0.10 (0.65-0.94). Finally, the correlation of VD value between the CM and SA methods in all subjects was 0.43 ± 0.33 (0.97-0.00).

Table 1.

The results of r-squared fitting curves all subjects

Subject	CM	SA
1	0.67 ±0.12	0.67 ±0.12
2	0.60 ±0.11	0.62 ±0.11
3	0.79 ±0.09	0.79 ±0.08
4	0.68 ±0.12	0.69 ±0.12
5	0.58 ±0.12	0.60 ±0.12
6	0.81 ±0.07	0.81 ±0.09
7	0.89 ±0.06	0.89 ±0.07
8	0.58 ±0.12	0.61 ±0.11
9	0.62 ±0.12	0.63 ±0.12

Fig. 4.

Comparison of the net uptake ratio K_i, K₁, and VD estimated by CM and SA in the nine subjects at 180 min. The last column shows the TBR images. CM image appears smooth with thick definition of the tumor

Fig. 5.

Boxplot comparing K_i, K₁, and VD among CM and SA and TBR in the whole tumor volume. The values were from the 9 subjects and were normalized to the mean in each group

On the contrary, the images accounting for the accumulation compartments calculated with (2) for CM and with (4) for SA showed a clear tumor definition and smooth images on SA than on CM (Fig. 6). The values of B₁ and B_lim are indicated in each image of SA. The input response functions of the accumulation component of the tumor from CM and SA are plotted in Fig. 7 for the three subjects, depicting comparable behavior.

Fig. 6.

CM accumulation compartment versus SA accumulation component in the nine subjects. Tumor images with SA appear better defined

Fig. 7.

Input response functions of the tumor from CM and SA as respectively defined in (2) and (4). a-i represents the nine subjects, respectively

In Fig. 8, the CM parametric images of K₁ and k₃ in all subjects were compared showing a mismatch of k₃ in the tumor of subjects 1, 4, 5, 6, and 9, while k₃ matched the tumor for other subjects. This behavior would be depending on the shape of TAC as in Fig. 3. The K₁ image, which is the perfusion, partly reproduced the shape of the tumor as in the measured image except for some voxels appearing relatively with low intensity. For subjects 1 and 7, the k₃ image showed an accumulation of 18F-FMISO in the necrosis region which was not shown in any other image from the 7 subjects. On the other hand, comparing of measured with K_i images in Fig. 4 shows that the high uptake voxels were matched between K_i and the measured images.

Fig. 8.

Comparison of CM K₁ and k₃ images with the measured images taken at 180 min in the nine subjects

Discussions

Hypoxia is a significant characteristic of solid tumors in most cancers [18]. In GBM tumors, hypoxia does not only contribute to tumor resistance but also has a role in the tumor spreading [19]. In this study, we showed the temporal variations of 18F-FMISO uptake within the tumor volume. In most clinical studies, the protocol utilized is a late static 18F-FMISO image taken between 2 and 4 h post-injection due to the slow kinetics of 18F-FMISO [8]. In a previous work of Thorwarth et al., they found that some regions of the tumor with high uptake of 18F-FMISO can decrease in intensity after 120 min post-injection in comparison to other regions of the same tumor in the same subject, and thus, it is not recommended to rely on image intensity at 120 min post-injection to determine hypoxia [11].

Parametric images were able to identify the perfusion and the retention of 18F-FMISO within the tumor volume. The TBR images showed a high uptake of 18F-FMISO in some regions not matching the k₃ parametric images. The net uptake rate K_i images from the spectral analysis were used to estimate the accumulation inside the tumor. Regions of high uptake in the measured image do not always show high uptake in the K₁ images. This high uptake of tracer in the measured images could be explained as a mix of tissue with good perfusion and diffusion-limited hypoxia [11]. In addition, the degree of necrosis inside the tumor can affect the tracer uptake since 18F-FMISO molecules do not bind to the necrotic regions, but they can freely diffuse inside this region [20]. The new finding from this study was the generation of the 18F-FMISO accumulation component with spectral analysis as a parametric image of hypoxia. This image can be further segmented where the voxels of higher intensity can be subjected to radiation boosts in radiotherapy.

This study may have some limitations. The noise in TACs data can affect both CM and SA when applied at the voxel basis. However, the values of kinetic parameters in both methods have different susceptibility to noise effect. K_i in CM is less sensitive to noise than k₃ [21]. In SA, K₁ is estimated as the summation of amplitudes along all frequencies, while K_i is the amplitude at B = 0 which makes K_i more sensitive to noisy data. The main differences between CM-Ki and SA-Ki were the effect of noise in determining the accumulation exponential. This problem appeared at the beginning of the SA application, where the noise can influence exponential detection. Therefore, shifting of components from their true values were mainly affected the accumulation component, which assumed to have only one component in SA estimation. The restricted number of components and the relative calculation of K_i in CM minimize the effect of noise.

Conclusions

A dynamic acquisition of 18F-FMISO adds more accuracy and flexibility in locating the hypoxia regions in glioblastoma multiforme tumors. Its disadvantage is the time-consuming in the clinic which prevents from scanning more patients. The use of the compartmental modeling approach was shown to correlate with results from spectral analysis. The recourse to parametric images from different components and rate constants increases the accuracy of hypoxia determination. From this study, the accumulative 18F-FMISO component defined with spectral analysis was shown to be the most representative of hypoxia image based on the input response functions of the pixels, in addition to the enhanced contrast and smoothness of the image. The parametric k₃ image specifically showed a different pattern when the tumor has a continuous accumulation of 18F-FMISO.

Compliance with Ethical Standards

Conflict of Interest

Redha-alla Abdo, Frédéric Lamare, Philippe Fernandez, and M’hamed Bentourkia declare no conflict of interest.

Ethical Statement

All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the Helsinki Declaration as revised in 2013 and its later amendments or comparable ethical standards.

Informed Consent

Informed consent was obtained from all individual participants included in the study at the Université de Bordeaux-II.