Combination of pre-treatment dynamic [¹⁸F]FET PET radiomics and conventional clinical parameters for the survival stratification in patients with IDH-wildtype glioblastoma

The inclusion of mandatory molecular markers for diagnosis in the World Health Organization (WHO) Classification of Tumors of the Central Nervous System (CNS) in 2016 and revised in 2021 has led to a more rigid definition of prognostically distinct entities [1, 2]. In particular, the isocitrate dehydrogenase (IDH)-wildtype status is associated with a worse prognosis in adult diffuse astrocytic gliomas [3] and results in the diagnosis of a glioblastoma, WHO grade 4, according to the 2021 WHO classification. Additional predictive markers such as the methylation status of the O-6-methylguanine-DNA-methyltransferase (MGMT) promotor further help to stratify brain tumor patients according to their individual risk profile [4]. However, even within the distinct molecularly defined tumor type of IDH-wildtype glioblastomas, few patients survive several years whereas others remain short-term survivors (STS) and decease within the first year, indicating further potential for improvement regarding patient stratification [5]. Balancing aggressive treatment including radiation and chemotherapy with quality of life is critical for patients [6].Therefore, additional prognostic markers beyond established molecular genetic markers and a stratification of survival beyond the neuropathological classification of brain tumors would be helpful to further improve individual prognostication and guide patient management accordingly.

Molecular imaging using positron emission tomography (PET) with radiolabeled amino acids such as O-(2-[¹⁸F]-fluoroethyl)-L-tyrosine ([¹⁸F]FET) has been applied successfully for the characterization and evaluation of primary brain neoplasms [7‐9]. Hence, PET imaging was recommended by the Response Assessment in Neuro-Oncology (RANO) Working Group as useful imaging method in addition to conventional magnetic resonance imaging (MRI) in the clinical management of brain tumor patients [10]. Especially dynamic [¹⁸F]FET PET has been shown to be helpful for non-invasive tumor classification [11] and for individual prognostication even within defined molecular subgroups [7, 12]. Here, radiomics have recently gained increasing interest as a promising non-invasive tool, where quantitative features are extracted from medical images and combined with clinical and genomic information to establish predictive models [13, 14]. However, up to now, there is no radiomic approach based on dynamic [¹⁸F]FET PET data which aims to perform survival stratification specifically in patients with an IDH-wildtype glioblastoma, despite being one of the most common and aggressive brain tumors.

Therefore, the purpose of this study was to build and evaluate a prediction model, which incorporates clinical parameters and radiomic features extracted from static as well as dynamic [¹⁸F]FET PET for an individualized survival stratification in patients with a newly diagnosed IDH-wildtype glioblastoma.

This study illustrates that integration of radiomics based on dynamic [¹⁸F]FET PET may improve the assessment of short-term survival probability in patients with newly diagnosed IDH-wildtype glioblastoma. As opposed to prediction models based on clinical parameters or radiomic features alone, specifically a combined clinical-TTP model including both clinical parameters and an additional radiomic signature derived from dynamic PET accomplished a higher prognostic value for short-term survival.

Several studies have analyzed the role of [¹⁸F]FET PET for the assessment of survival probability in patients with glioma [7, 8, 12, 32‐35]. It has been reported that a large biological tumor volume (BTV) on static [¹⁸F]FET PET [32, 33, 35] as well as a short TTP_min extracted from dynamic [¹⁸F]FET PET at initial diagnosis are associated with STS [7, 12, 34, 35]. Besides, Bauer et al. showed that TTP_min is an independent prognostic factor for overall survival, reaffirming the value of dynamic [¹⁸F]FET PET in the prediction of survival in glioma patients. Yet, initial radiomics data in high-grade glioma have been provided by MRI studies, achieving high AUC values for the prognostication of overall survival in the range of 0.652–0.858 in the test cohort [36‐40] demonstrating that radiomics might be a valuable tool to estimate survival in brain tumor patients. Meanwhile, first promising studies have brought [¹⁸F]FET PET–based radiomics into the focus: Radiomic features extracted from static [¹⁸F]FET PET showed better accuracy than conventional static parameters (e.g., TBR_max) to identify pseudoprogression [13]. For the differentiation between radiation injury and recurrence of brain metastasis, textural features extracted from [¹⁸F]FET PET had a diagnostic accuracy of 83% [41]. Carles et al. reported that [¹⁸F]FET PET radiomics could contribute to the prognostic assessment [42], and Paprottka et al. established a promising tool for objective differentiation of tumor progression from treatment-related changes by combining [¹⁸F]FET PET and multiparametric MRI [43]. However, those initial studies only analyzed static [¹⁸F]FET PET features without taking into account important clinical parameters and, furthermore, no study so far has utilized dynamic [¹⁸F]FET PET–based radiomics to assess the probability of poor prognosis within distinct molecular brain tumor types.

The present study used clinical parameters combined with [¹⁸F]FET PET radiomic features to develop combined clinical-radiomic models. A model based on clinical data only, built from six important survival-related clinical parameters, achieved an AUC of 0.69 in the independent testing cohort. A TBR model, built from two static [¹⁸F]FET PET features, achieved an AUC of 0.63 in the testing cohort and thus did not perform better than the clinical model. The TTP model, however, generated from six dynamic [¹⁸F]FET PET features, achieved an AUC of 0.71 in the testing cohort, thus slightly exceeding the clinical-only model and outranging the TBR-only model, highlighting the importance of dynamic PET data in the context of survival-related analyses. The combined purely imaging-based TBR-TTP model achieved only slightly better results than each model alone (AUC of 0.74 vs. AUC of 0.63 and AUC of 0.71). Eventually, the merger of the TTP radiomic signature and clinical data, resulting in the combined clinical-TTP model, achieved best predictive performance with an AUC of 0.74. Integrated discrimination improvement (IDI) was calculated between the clinical model and the combined clinical-TTP model [44]. The value of IDI was 0.1089, which was greater than 0, and the P value was 0.023, which was statistically significant. It indicated that the combination of TTP radiomics and clinical data, compared to clinical parameters alone, led to an improved ability of the model to identify patients at risk. Although intriguing to speculate that the clinical-TBR-TTP model would achieve highest accuracy as it includes all available information, the AUC did not improve, which may be related to the limited value of TBR information in this context, but this should be re-evaluated in larger cohorts. Taken together, as previously shown for other entities, it seems beneficial not to narrow the view to the clinical information alone when constructing a predictive model but to include radiomic signatures in clinical prediction studies as well, as the combination of clinical and radiomic information seems to be of particular value with regard to survival risk prediction [45]. When considered on its own, an AUC of 0.74 still does not seem satisfactory, as further underscored by a positive predictive value for the identification of a short-term survivor of only 47.1% even for the best model (see Table 3). From a clinical point of view, the positive and negative predictive values are highly useful metrics in the context of decision-making as they give an estimate on the correctness of a prediction. In the clinical setting, it would be particularly beneficial to identify patients at risk for short-term survival in order to facilitate the selection of more aggressive treatments or earlier inclusion in experimental treatment studies, rather than just standard treatment, to which approximately 30% of patients do not respond. However, also the identification of long-term survivors would be helpful in the clinical routine, as pseudoprogression can occur in one-third of the patients and may, when misinterpreted as tumor progression on MRI, lead to a premature cessation of an effective treatment. Of note, while the positive predictive value was extremely low in all models, the negative predictive value, reflecting the predictability of long-term survival, reached 84% in the best model. Therefore, even though the overall accuracies of our prediction models may not yet be satisfactory for the clinical use and the low positive predictive values impede the prediction of a short-term survivor, the high negative predictive value may be helpful for clinical decision-making. Our study supports that within a neuropathologically homogenous group of aggressive IDH-wildtype glioblastomas, especially the combination of different types of information (in this case clinical data and radiomic signature) can add value to a survival prediction model and consequently hints to the potential, which lies in the inclusion of even further image-based information. Indeed, one might speculate that the addition of conventional MRI data and in a next step more sophisticated MRI data such as perfusion or diffusion-weighted MRI may further increase the power of survival risk prediction of the combined clinical-TTP model [46], but such analyses require a standardized imaging protocol to assure comparability of MRI-based radiomic features. In other tumor entities as well, especially multiparametric imaging approaches have shown highly promising results for survival prediction, e.g., reaching an accuracy of up to 98% in a study on cervical cancer as compared to only 56–60% for prediction models using the standard clinical variables alone [47, 48]. Accordingly, dual PET imaging studies including other tracers than [¹⁸F]FET in IDH-wildtype glioblastoma, such as TSPO-ligands which offer complementary information to the [¹⁸F]FET uptake [49], are of high potential to further increase the power of survival prediction models, as exemplified by recent successful multi-tracer PET prediction approaches in other entities, such as prostate cancer [50].

Although the number of patients included in the current study is by far higher than in most previous [¹⁸F]FET PET radiomics studies, a further increase in patient numbers may in future result in outperforming radiomics-only based approaches, as already shown in large-scale analyses for other medical settings [51]. According to the above-generated multivariate LR-based formulas, the known risk factors of high WHO grade, unmethylated MGMT promoter, TERTp mutation as well as higher patient age and lower KPS at diagnosis of IDH-wildtype glioblastoma were more likely associated with short-term survival [52‐55]. However, gender has different correlations in different formulas, which is inconsistent with the literature [53], although the weight of this parameter was low. This may likewise be due to the relatively low number of patients included in this study.

Whereas, in clinical routine, established dynamic [¹⁸F]FET PET parameters such as the time–activity curve and/or the slope are usually only derived from representative subvolumes of interest within the tumor [7, 9, 56], in the current study every single voxel of the tumor was analyzed in order to generate whole-tumor TTP maps of dynamic [¹⁸F]FET PET images. This comprehensive whole-tumor approach facilitated radiomic features extraction in dynamic image data and ensured to account for heterogeneity of uptake kinetics which has a major clinical impact when assessing brain tumors in dynamic [¹⁸F]FET PET [57]. In this context, a relationship between tumor heterogeneity and the STS group could be found in the feature ClusterProminence (CP). CP belongs to the Gray Level Co-occurrence Matrix (GLCM) and measures the skewness and asymmetry of the GLCM. A higher value implies more asymmetry while a lower value indicates a peak near the mean value and less variation around the mean. This correlation with the STS group indicates that a patient with a heterogeneous tumor in dynamic [¹⁸F]FET PET images is more likely to be identified as high-risk patient for short-term survival. Another exemplary radiomic feature, which is associated with the STS group, is Maximum 3D diameter, 3D shape feature. The latter is defined as the largest pairwise Euclidean distance between tumor surface mesh vertices. This correlation, in simplified terms, indicates that patients belonging to the STS group have a tumor that shows large spread on PET. This finding is consistent with the literature—large tumor volumes on [¹⁸F]FET PET were reported to be associated with poor overall survival in glioblastoma patients before radiation therapy with concomitant and adjuvant temozolomide [32, 33]. Details of other features are shown in the Supplementary information.

There are several limitations to this study. Only single-center data have been investigated, which led to the relatively small sample size and the lack of external validation. Yet, only single-center data have been chosen in this study since dynamic [¹⁸F]FET PET is not always acquired routinely in other centers and pooling PET data with differences in time framing, image reconstruction algorithm, and scanner type may require prior implementation and validation of, e.g., feature harmonization procedures [58]. Moreover, it should be noted that almost all previous [¹⁸F]FET PET radiomics studies have been performed with much smaller numbers of cases. The reliability of the reported scores was additionally evaluated using nested cross-validation [59] with five random splits in the outer loop, yielding a high AUC variability of 10% for the TTP model, 15% for the TBR model, and 11% for the clinical model (Supplementary material S4). Thereby, different radiomic signatures were obtained for each split of the outer loop since feature selection and model building are not robust when dealing with small sample sizes. Feature selection represents a challenge and has an impact on the performance of prediction models. Other feature selection methods comprise, e.g., filter methods such as minimum redundancy maximum relevance (MRMR) or ensemble methods, which provide a good balance between robust feature selection and model performance. Wrapper methods such as RFE have the advantage that feature dependencies can be modeled and that they interact with the classifier, while also bearing the risk of overfitting [60]. To enable standardized segmentation of tumor regions, only positive [¹⁸F]FET PET images were included. Furthermore, MRI-based radiomics, as a more widely established and complementary tool, were not included in this study. Future studies may benefit from the combined use of multiparametric MRI data.

This study built and evaluated prediction models for survival combining both radiomic features extracted from static and dynamic [¹⁸F]FET PET and clinical parameters. Specifically, the combination of clinical parameters with radiomics based on dynamic [¹⁸F]FET PET data achieved a higher prognostic accuracy for the individualized assessment of short-term survival in patients with newly diagnosed IDH-wildtype glioblastoma in comparison to models using conventional clinical parameters only. Although the final accuracy remained moderate, the integration of dynamic [¹⁸F]FET PET radiomic data into clinical prediction models may improve patient stratification beyond established prognostic markers. Future prospective radiomic studies using multimodal imaging data are needed to evaluate whether the integration of additional imaging parameters may further improve the prognostic performance and enhance the clinical interpretation of the study results.