Background
Structural imaging with T
1-weighted magnetic resonance imaging (MRI) [
1], which is the gold standard in clinical glioma assessment, is restricted to the interpretation of properties like tumour contour, localisation, and enhancement pattern [
1]. Besides, several functional MRI techniques have shown relevance for prediction of malignant transformation, involving, e.g. perfusion-weighted imaging (PWI) yielding information on relative cerebral blood volume and flow (rCBV, rCBF) [
2‐
4]. In contrast, positron emission tomography (PET) with amino acids aims to directly image an elevated amino acid metabolism of rapidly proliferating tumour cells [
5‐
7]. According to the report on response assessment in neuro-oncology (RANO), dynamic O-(2-
18F-fluoroethyl)-
l-tyrosine (
18F-FET) PET has shown its usefulness in diagnosis, in prognosis of tumour progression, and in assessment of treatment response [
8].
The current standard procedure for retrieving information from dynamic
18F-FET PET consists of evaluating parameters such as the tumour-to-background ratio (TBR) at a certain time point, the late slope, the time-activity curve (TAC) pattern, and the time-to-peak (TTP) [
9‐
16]. In particular, the TTP and the TAC pattern have proven to be suitable for identification of tumour recurrence or progression [
12,
13,
17], and for glioma grading [
14,
15,
18]. Pharmacokinetic modelling of
18F-FET uptake has also been considered. However, to our knowledge, its clinical relevance could not be shown yet, and the requirement of (metabolite-corrected) plasma-input data impairs the clinical applicability [
19,
20]. While a slowly increasing TAC is characteristic of low-grade gliomas, the TAC of high-grade gliomas tends to exhibit a short TTP and decreasing TAC [
17,
21]. Those parameters are most frequently derived from a mean volume-of-interest (VOI)-TAC of the entire tumour or from the hot-spot of the tumour with a 90% isocontour [
17,
22]. However, in case of heterogeneous tumours, it may occur that the hot-spot in summation images does not correspond to the tumour fraction defined as most suspicious regarding tumour aggressiveness according to TTP and TAC pattern. This may potentially lead to an underestimation of malignancy and might impair treatment planning. Recent approaches in current research aiming to improve the assessment of tumour characteristics include, e.g. a slice-by-slice TAC analysis or the extraction of texture parameters from static
18F-FET PET images [
23,
24].
The goal of this study was to investigate the intra-tumoural distribution of the abovementioned diagnostically relevant kinetic and static parameters derived from dynamic 18F-FET PET data on a voxel basis. A comparison with VOI-based methods, as currently utilised for non-invasive glioma characterisation in clinical routine, is provided.
Discussion
In this study, we established an automated and reader-independent method for voxel-wise
18F-FET PET glioma analysis, which enables a fast identification of sub-volumes consisting of voxels with aggressive high-grade kinetics. By quantifying the intra-tumoural parameter distribution with percentage volume histograms, we found significant differences between WHO grades and between molecular genetic groups. Both, association with WHO grade and
IDH mutation status, were higher for PVH data compared to VOI-based parameters in most cases. Interestingly, sub-group analyses showed that in the special case of
IDH-wt gliomas, the fraction with early peak or negative slope was significantly higher in WHO grade III compared to WHO grade IV gliomas, with simultaneously significantly higher PVH
TBR,20–40 > 2 in WHO grade IV gliomas. Aggressive sub-volumes defined by TTP < 20 min p.i. and negative Slope
15–40 showed high overlap with each other, but a low overlap with TBR
5–15 > 2- and TBR
20–40 > 2-defined hotspots, indicating a possible complementarity of the investigated kinetic and static parameters. The corresponding parametric images as presented in Fig.
5 may provide valuable information for a fast visual screening of glioma tissue. In summary, this study demonstrates the relevance and suitability of tumour heterogeneity assessment on a voxel basis with static and kinetic
18F-FET PET parameters for a differentiated characterisation of gliomas, although the clinical applicability of parametric 3D information yet requires a comprehensive validation by utilising stereotactic biopsies.
In this context, an elaborate understanding of the underlying processes of
18F-FET uptake is crucial and a matter of current research [
20,
31‐
35]. So far, various studies suggest that regional information from static
18F-FET PET images and from MR-based morphological and functional images is complementary, showing only moderate overlap and low spatial correlation [
36‐
39]. Still, tissue properties such as rCBV and rCBF might be relevant for the delivery of
18F-FET, potentially influencing
18F-FET uptake behaviour. rCBF was found to correlate significantly with early slope (0–5 min p.i.) in
18F-FET PET and with TBR (20–40 min p.i.), however, not with TAC patterns and late slope (10–50 min p.i.) [
40]. Recently, a negative correlation of rCBV and late slope (10–30 min p.i.) and a positive correlation with TBR (10–20 min p.i.) could be shown; however, only a small fraction of the variance of early and late FET uptake could be explained by rCBV [
38]. Therefore, it was concluded that rCBV and
18F-FET PET provide congruent and complementary information on the underlying processes. While late TBR may mainly reflect specific trapping within tumour cells, the early TBR and the TAC pattern may be influenced by rCBV and rCBF [
38,
41]. Correlation of
IDH mutation status with MRI parameters has among others shown that
IDH-wt gliomas tend to exhibit high rCBV values, which is a robust estimate of tumour angiogenesis [
32,
35]. In order to retrieve comprehensive information on the underlying processes and their influence on
18F-FET uptake, further investigations may combine information from PWI and pharmacokinetic modelling with dynamic
18F-FET PET data, also considering blocking studies.
Various studies were published evaluating thresholding techniques optimised for the reproduction of true object boundaries in PET images, possibly taking into account different image characteristics [
42‐
45]. The currently established method for BTV definition was verified with at least one biopsy per patient, which was utilised for an optimisation of sensitivity and specificity and resulted in the optimal TBR cut-off of 1.6 [
15,
27]. As shown previously in mice, a threshold relying on background and maximal uptake within the tumour is superior for reproduction of histologically proven glioma boundaries [
46]. Hence, future studies considering glioma segmentation in humans, possibly further including information from the characteristic kinetics of the different glioma types, are desirable.
The proposed voxel-wise analysis including TTP and Slope
15–40 maps and percentage volume histograms of static and kinetic parameters has the potential to provide encompassing information not only for planning of biopsy, surgery, or radiation therapy but also for prognosis, follow-up, and prediction of tumour recurrence based on improved 3D information regarding the local aggressiveness of tumour tissue. In this context, this study has two limitations which need to be addressed in future studies. Firstly, this work would benefit from a correlation analysis of histopathologically assessed tumour heterogeneity and the tumour heterogeneity indicated by the proposed parametric 3D maps. Secondly, voxel-TACs are prone to noise in dynamic PET data, especially for shorter time frames. In this study, sensitive parameters TTP and Slope
15–40 were derived directly from single-voxel TACs without the application of TAC smoothing or fitting in order to avoid the introduction of bias, i.e. change in temporal pattern, from TAC pre-processing, and allow for an easy adoption by other research centres. An exemplary simple method for per-frame noise suppression with a spatial Gaussian filter was included and showed that PVH data changed while the ability to differentiate glioma types was preserved, which further underlines the need for stereotactic biopsies. Although the incorporation of a kinetic model which is suitable to describe
18F-FET pharmacokinetics seems conceivable, provided that appropriate blood input data are available, voxel-based fitting of complex models might also be sensitive to noise [
19].
The presented data indicate the direct applicability for non-invasive glioma grading and prediction of molecular genetic profile. This is important, since the WHO classification was updated [
26], and stratification is now based on molecular genetic information, i.e.
IDH-wt gliomas are considered as having the same prognosis as glioblastomas themselves. A direct application is the clinical assessment of lesions suspected of glioma, in particular for the selection of the subsequent clinical steps such as biopsy, resection, or “watch and wait”, but also for risk-stratification in non-contrast-enhancing gliomas (
IDH-mut vs.
IDH-wt). The next steps may further include multi-parametric 3D analysis, machine learning approaches, the evaluation of the influence of small scale motion on voxel-wise analysis, and the assessment of the robustness of alternative methods for the voxel-wise characterisation of gliomas, such as pharmacokinetic modelling or the inclusion of information from other imaging modalities like perfusion-weighted imaging.