Prediction of TERTp-mutation status in IDH-wildtype high-grade gliomas using pre-treatment dynamic [18F]FET PET radiomics

Purpose To evaluate radiomic features extracted from standard static images (20–40 min p.i.), early summation images (5–15 min p.i.), and dynamic [18F]FET PET images for the prediction of TERTp-mutation status in patients with IDH-wildtype high-grade glioma. Methods A total of 159 patients (median age 60.2 years, range 19–82 years) with newly diagnosed IDH-wildtype diffuse astrocytic glioma (WHO grade III or IV) and dynamic [18F]FET PET prior to surgical intervention were enrolled and divided into a training (n = 112) and a testing cohort (n = 47) randomly. First-order, shape, and texture radiomic features were extracted from standard static (20–40 min summation images; TBR20–40), early static (5–15 min summation images; TBR5–15), and dynamic (time-to-peak; TTP) images, respectively. Recursive feature elimination was used for feature selection by 10-fold cross-validation in the training cohort after normalization, and logistic regression models were generated using the radiomic features extracted from each image to differentiate TERTp-mutation status. The areas under the ROC curve (AUC), accuracy, sensitivity, specificity, and positive and negative predictive value were calculated to illustrate diagnostic power in both the training and testing cohort. Results The TTP model comprised nine selected features and achieved highest predictability of TERTp-mutation with an AUC of 0.82 (95% confidence interval 0.71–0.92) and sensitivity of 92.1% in the independent testing cohort. Weak predictive capability was obtained in the TBR5–15 model, with an AUC of 0.61 (95% CI 0.42–0.80) in the testing cohort, while no predictive power was observed in the TBR20–40 model. Conclusions Radiomics based on TTP images extracted from dynamic [18F]FET PET can predict the TERTp-mutation status of IDH-wildtype diffuse astrocytic high-grade gliomas with high accuracy preoperatively. Supplementary Information The online version contains supplementary material available at 10.1007/s00259-021-05526-6.


Introduction
Mutations in the telomerase reverse transcriptase promoter (TERTp), leading to telomerase activation and lengthened telomeres, play an important role in the formation of brain cancer and individual prognosis [1][2][3]. In diffuse astrocytic high-grade gliomas without mutation of the isocitrate dehydrogenase gene (IDH-wildtype), TERTp mutations are reported to be associated with poor overall survival [4][5][6]. Molecular genetic analysis of the TERTp-mutation status has therefore gained increasing attention in the clinical routine diagnosis of IDH-wildtype diffuse astrocytic gliomas and will be included in the upcoming glioma WHO classification [7][8][9].
Molecular imaging using positron emission tomography (PET) with radiolabelled amino acids such as O- (2-[ 18 F]fluoroethyl)-L-tyrosine ([ 18 F]FET) is a useful tool for the characterization and evaluation of primary brain neoplasms [10][11][12], and its application in the clinical management of brain tumour patients has been recommended by the Response Assessment in Neuro-Oncology (RANO) Working Group [13][14][15][16][17]. While static image data (standard 20-40 min summation images) are particularly used for the delineation of the tumour extent, the assessment of dynamic [ 18 F]FET PET data has been shown to provide additional information about tumour biology [18]. More aggressive gliomas (i.e. high-grade gliomas and/or IDH-wildtype gliomas) were shown to be characterized by a high tracer uptake within the first 5-15 min post injection (p.i.) with subsequent curve decrease, while less aggressive gliomas (i.e. low grade gliomas and/or IDH-mutant gliomas) typically show a slowly increasing [ 18 F]FET uptake with highest values in the later time frames [12,19,20]. As the early peak uptake in aggressive gliomas is missed in the standard 20-40 min p.i. summation images, it does not surprise that the maximal tumour-to-background ratio (TBR max ) evaluation obtained in early summation images (5-15 min p.i.) was reported to perform better than the standard static TBR max values (20-40 min p.i.) for the differentiation between low-grade and high-grade gliomas [17], which led to the suggestion to include these early summation images for a better glioma characterization. Another interesting parameter derived from dynamic [ 18 F]FET PET is the minimal time-to-peak (TTP min ), which is extracted from the time-activity-curves and was reported to provide prognostic information [21]. Interestingly, an early TTP min was associated with an aggressive disease course in newly diagnosed gliomas and was able to predict an IDH-wildtype status [22,23]. Yet, in our recently published study investigating [ 18 F]FET uptake characteristics in TERTp mutant and TERTp wildtype glioblastomas, neither the standard TBR max as static parameter nor TTP min as dynamic parameter were associated with the TERTp-mutation status [24].
In recent years, radiomics have been increasingly investigated as a promising non-invasive tool for accurate diagnosis and prognosis assessment by converting medical images into high-dimensional quantitative image features and establishing predictive models [25][26][27][28][29][30][31][32]. However, radiomics have not been applied for the detection of TERTp mutations on [ 18 F] FET PET images so far. Therefore, the aim of this study was to evaluate radiomic features extracted from standard static images (20-40 min p.i.), early summation images (5-15 min p.i.) as well as dynamic [ 18 F]FET PET images for the prediction of the TERTp-mutation status in patients with newly diagnosed IDH-wildtype diffuse astrocytic highgrade glioma. F]FET-negative gliomas (tumourto-background ratio, TBR < 1.6) were excluded. All patients had given written informed consent prior to the PET scan as part of the clinical routine. The retrospective analysis of PET imaging data was approved by the institutional ethics committee (604-16). A total of 61% of the investigated patients (97/159) have been evaluated in a previous study [24].

Histopathology and molecular genetic analysis
Histopathology and molecular genetic analyses were performed at the Institute of Neuropathology, LMU Munich, Germany. All patients initially classified according to the 2007 WHO brain tumour classification [34] were reclassified according to the 2016 WHO classification [33]. The IDH-mutation status and TERTp-mutation status were evaluated according to clinical standard protocols [35,36]. [ 18 F]FET PET scans were performed at the Department of Nuclear Medicine, LMU Munich, Germany. Images were acquired by using an ECAT EXACT HR + PET scanner (Siemens Healthineers, Inc., Erlangen, Germany) with the standard protocol [11,37]. Exactly 180 MBq of [ 18 F]FET were injected after a 15-min transmission scan with a 68 Ge rotating rod source. After tracer injection up to 40 min post injection in 3-D mode consisting of 16 frames (7 × 10 s, 3 × 30 s, 1 × 2 min, 3 × 5 min, and 2 × 10 min) with a reconstructed voxel size of 2.03 × 2.03 × 2.43 mm 3 and matrix size of 128 × 128 × 63, dynamic emission recording was finished. Two-dimensional filtered back-projection reconstruction algorithm using a 4.9-mm Hann Filter was applied for image reconstruction, then corrected for attenuation, decay, dead time, and random and scattered coincidences. When relevant motion was visible in dynamic PET data, a frame-wise correction was performed by using PMOD fusion tool (version 3.5, PMOD Technologies, Zurich, Switzerland) after framewise checking for motion.

Segmentation of tumour volumes and brain background
First, a background activity was extracted from a large crescent-shaped volume of interest (VOI) in the contralateral healthy hemisphere as published previously [38]. For tumour segmentation, a VOI was drawn using a TBR-threshold of 1.6 in static 20-40 min p.i. summation images as suggested by Pauleit et al. [39]. All segmentations were processed within the PMOD View tool (version 3.5, PMOD Technologies, Zurich, Switzerland).

Feature selection
Before feature extraction, a stratified random split was used to assign 70% of the patients to the training cohort (n = 112) and the remaining 30% to the testing cohort (n = 47), with a balanced distribution of TERTp-wildtype and TERTp-mutation.
Features were standardized as follows: for each feature, we calculated the mean value and the standard deviation. The mean value was subtracted from each individual value, which was then divided by the standard deviation. Feature normalization was computed only in the training cohort and then applied on the testing cohort. Since the number of features was large, we compared the similarity of each feature pair. If the Pearson correlation coefficient (PCC) value of the feature pair was larger than 0.99, we removed one of them. After this process, the number of the features was reduced and each feature was independent to each other. The recursive feature elimination (RFE) based on logistic regression classifier was performed to reduce redundant features and select potential TERTp-mutation related features [45]. Considering the imbalance of comparison groups, we performed the weighted logistic regression in the 'balanced' mode, which gives higher weight to the minority class and lower weight to the majority class and therefore automatically adjusts weights inversely proportional to class frequencies in the input data [46]. Each iteration removes a feature which is considered least important. After stratified splitbased 10-fold cross-validation, the area under the receiver operating characteristic curve (AUC) of the model in the training cohort was used to determine the optimal number of features.

Model construction and testing
Logistic regression (LR) models were built to predict the TERTp-mutation status by fitting the selected radiomic features. Each model was generated by using only the radiomic 1 3 features extracted from each image (i.e. TBR [5][6][7][8][9][10][11][12][13][14][15] , TBR  , and TTP images) separately. According to the coefficients of selected features generated by the LR models [47], the risk probability of TERTp-mutation was calculated by the following formula: x is the value of selected features, is the coefficient of selected features, and 0 represents the intercept. In case of P > 0.5 , TERTp-mutation status was considered as positive by the LR model.
Model testing was applied to the independent testing cohort, which was not involved in the process of model training. The workflow of the process is presented in Fig. 1.

Statistical analysis
To evaluate the model performance, receiver operating characteristic curve (ROC) analysis was performed in the training and testing cohort. The AUC was calculated as quantitative measure to illustrate diagnostic power. The accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were calculated. 95% confidence intervals (CI) were calculated by using a non-parametric bootstrap method, which was repeated 1000 times to get a bootstrap distribution of the results.
Categorical variables or continuous variables were reported as numbers and percentages or as mean and standard deviation. Categorical variables were compared by the P(y = 1|x; ) = 1 1 + e − T x χ 2 test, and continuous variables were compared by the Mann-Whitney U test. P < 0.05 were considered statistically significant. Statistical analyses were programmed in Python (v. 3.8.5; https:// www. python. org/).

Patient characteristics
A total of 159 patients (median age, 60.2 years; range, 19-82 years) were enrolled in this study. Exactly 31 patients (19.50%) were diagnosed with TERTp-wildtype, and 128 patients had TERTp mutation. The clinical characteristics are presented in Table 1. There were no significant differences between the training and testing cohorts with regard to age, sex, WHO grade, and TERTp mutation status, with TERTp-wildtype rates of 19.64% and 19.15%, respectively.

Radiomic feature extraction and selection
In this study, 107 radiomic features of candidates were gen-   Fig. 2 The feature selection process of the RFE method. Each iteration removes a feature that is considered least important and corresponds to a 10-fold cross-validation. After 10-fold cross-validation, the AUC of the model in the training cohort was used to determine the optimal number of features. The minimum AUC of feature num-ber was selected. a TBR 5-15 model, b TBR  , and c TTP model; 9, 14, and 10 features were selected respectively. RFE recursive feature elimination, AUC area under the receiver operating characteristic curve

Diagnostic Validation of the TBR 20-40 model, TBR 5-15 model, and TTP model
According to the above-mentioned formula, the risk probabilities of TERTp-mutation were calculated. The coefficients of selected features in the TBR 20-40 model and TBR 5-15 model are shown in Table S1. The coefficients of selected features in the TTP model are shown in Table 2.
Detailed information about the performance of each model is shown in Table 3.

Discussion
Our study showed that radiomics based on dynamic [ 18 F] FET PET data can reliably predict the TERTp-mutation status of IDH-wildtype diffuse astrocytic high-grade gliomas. Best predictability was reached using the TTP model derived from dynamic PET, and weak predictive capability was obtained with radiomics based on early summation images (5-15 min p.i.), while no reliable information about the TERTp-mutation status was possible based on the standard summation images (20-40 min p.i.).
Previous studies have shown that patients with IDHwildtype TERTp-mutant glioblastoma have a significantly shorter progression free and overall survival compared to those with TERT-wildtype status. Therefore, TERTp-mutation status is now considered to be an important diagnostic and prognostic factor in primary glioblastomas and especially in patients with IDH-wildtype glioma [3,5,8,9,48]. TERTp-mutations indicate tumours that require aggressive and immediate treatments [3]. Hence, a preoperative tool for the prediction of a TERTp-mutation would be useful for early decision making and clinical management of patients with suspected glioma.
Several studies have analyzed the value of MRI based radiomics to predict the TERTp-mutation status in brain tumour patients [49][50][51]. Although these studies reported to achieve high accuracy values in the range of 79.88-93.80%, only WHO grade II or/and III gliomas have been considered and a limited number of patients has been investigated [49][50][51]. Besides, Tian et al. established a multiparameter MRI based radiomics model for the prediction of the TERTp-mutation status in patients with high-grade glioma [52], but ignored that TERTp-mutations play different roles in different IDH phenotypes [48].
Compared with conventional MRI, amino acid PET has been shown to be more sensitive in the definition of brain tumour extent [39], and dynamic [ 18 F]FET uptake parameters extracted from the TAC have shown to be an independent biomarker for prognosis [53,54]. Several studies have reported the informative value of [ 18 F]FET PET-based radiomics in personalized clinical decisions and individualized treatment selection [27][28][29]55]. Lohmann et al. found textural feature analysis in combination with TBRs to better differentiate brain metastasis recurrence from radiation injury than TBRs alone, and [ 18 F]FET PET radiomics achieved a higher accuracy than the best standard FET PET parameter (TBR max ) to diagnose patients with pseudoprogression [27,55]. Haubold et al. utilized multiparametric [ 18 F]FET PET/MRI and MR fingerprinting to decode and phenotype cerebral gliomas, which may serve as an alternative to invasive tissue characterization [28]. In addition, Carles et al. evaluated the prognostic value of [ 18 F]FET PET radiomics after re-irradiation, and found it could contribute to the selection of recurrent glioblastoma patients benefiting from re-irradiation [29]. However, all studies included radiomics based on standard static images (20-40 min p.i.) only and did not extract radiomic features derived from dynamic [ 18 F] FET PET as well as early summation images (5-15 min p.i.) even though two studies have shown the impact of dynamic parameters on radiomics [32,56]. Furthermore, no study has evaluated the potential to predict the TERTp-mutation status by [ 18 F]FET PET radiomics so far.
This study included standard static images (20-40 min p.i.), early summation images (5-15 min p.i.), and dynamic [ 18 F]FET PET images to develop the radiomic models. A total of 107 features were extracted from each  [24]. Interestingly, radiomics based on the standard TBR 20-40 model showed a low performance for the prediction of the TERTp-mutation status, and even the TBR 5-15 model, generated from nine early summation [ 18 F]FET PET features,  had an accuracy of only 66% and an AUC of 0.61 in the testing cohort. With a high prediction accuracy of 83% in the TTP model, our study demonstrates that radiomic features extracted from dynamic PET data can achieve a higher performance level than models based on static PET data. Remarkably, the sensitivity of the TTP model reached 92.1% in the testing cohort, so that patients with aggressive TERTp-mutant glioma can be identified non-invasively with high probability [3]. With the generated multivariate LRbased formula, health practitioners will be able to calculate the patient individual risk probability of bearing a TERTpmutation before neurosurgical intervention. Our study shows that even sophisticated radiomic analysis of static [ 18 F]FET PET imaging cannot replace dynamic acquisitions, at least with regard to the prediction of the TERTp-mutation status. Traditional dynamic [ 18 F]FET PET parameters such as the classification of the time-activity curve (increasing vs. decreasing or increasing vs. plateau vs. decreasing), the slope or the TTP min were most frequently calculated from a mean VOI-TAC of the tumour or from the hot-spot of the tumour with a 90% isocontour [10,12,19]. Considering the heterogeneity of gliomas, it may happen that the hotspot in standard summation images does not correspond to the most suspicious tumour aggressiveness when only considering TTP min and TAC and that, therefore, the most aggressive areas are inadvertently not evaluated. In contrast, we extracted the dynamic [ 18 F]FET uptake information in every voxel within the tumour VOI and generated TTP images. This approach, which was first introduced by Kaiser et al. [40,42], ensures that the dynamic information including the heterogeneity of uptake kinetics is extracted and that radiomics can be performed on the prognostically valuable dynamic data. The correlation between tumour heterogeneity and TERTp-mutation status can be considered in GreyLevelNonUniformityNormalized (GLNN) feature, which was used in the TTP model (see Table 2). GLNN belongs to Gray Level Dependence Matrix (GLDM), which is mathematically equal to first order-uniformity and is a measure of the homogeneity of the image array. A low value implies a greater heterogeneity, which was correlated with the TERTp-mutation, indicating that tumours with more heterogeneous TTP images are more likely to be classified as TERTp-mutant glioma.
Several limitations of this study should be discussed. First, the number of investigated patients is relatively small. However, it needs to be considered that we analyzed a very homogeneous group of patients with newly diagnosed and untreated IDH-wildtype diffuse astrocytic high-grade glioma. To exclude any influence by scanner type, all images in this study were derived from the same PET scanner, which limited the number of patients as well. In order to increase the number of patients, multi-centre validation studies are needed which, however, require phantom studies and harmonization of reconstruction parameters to make images from different PET scanners comparable. Another approach to directly harmonize features extracted from different devices may be to use the ComBat method [57]. In addition, our results are difficult to extrapolate to other centres, as the PET images analyzed in this study were acquired with our old PET scanner with fixed time frames, resulting in relatively long time frames (predominantly 5 and 10 min) in the dynamic analysis which could not be changed afterwards, and were reconstructed using filtered back-projection, while most PET centres now use other reconstruction methods such as ordered subset expectation maximization (OSEM). Furthermore, radiomic features were only extracted from the [ 18 F]FET-positive tumour VOI to construct the model. Besides the tumour VOI, the remaining image (with normal seeming tissue) may still contain invisible but useful information. To analyze the entire images, deep learning methods will be necessary. Furthermore, our study focused on PET-based radiomics only. A combination with MRI may improve the performance of the prediction model and should be evaluated in future studies.

Conclusion
While conventional [ 18 F]FET PET parameters assessed by standard analyses have previously shown no association with the TERTp-mutation status, radiomic models can predict the TERTp-mutation status of IDH-wildtype diffuse astrocytic high-grade gliomas with high accuracy preoperatively. Notably, this is only the case for radiomics based on dynamic image data (TTP model) instead of standard summation images (20-40 min). Further external validation in multicentre studies with a larger number of patients is needed to evaluate the potential for clinical applications.