Imaging-genomics reveals driving pathways of MRI derived volumetric tumor phenotype features in Glioblastoma

Glioblastoma (GBM) is a highly invasive and diffuse WHO grade IV tumor and is the most lethal central nervous system malignancy with an annual age-adjusted incidence rate of 3.19/100,000 per population [1]. Despite aggressive surgical therapy, radiation therapy, and temozolomide administration the 2-year survival rate remains around 27 % [2]. As a result, recent investigations have focused on capitalizing on the high molecular heterogeneity of gliomas to develop personalized treatment strategies [3].

One promising avenue of these investigations involves quantitative analyses of radiographic data, where imaging modalities are used to quantify tumor phenotype noninvasively. In magnetic resonance imaging (MRI), GBM tumors exhibit strong phenotypic features such as Necrosis, Edema, Contrast Enhancement, and Tumor Bulk (Fig. 1). These properties can be captured without and with intravenous administration of gadolinium-based contrast agents including T1-weighted or FLuid-Attenuated Inversion Recovery (FLAIR) (Fig. 2). In this way, visible tumor phenotype features can be systematically quantified.

As the underlying drivers of these phenotypes are biological in nature, recent efforts have been conducted indicating underlying genetic characteristics of imaging features. For example, tumor “Ring Enhancement” was found to be significantly associated with unmethylated MGMT promoter status [4, 5], which is known to be a biomarker for response to temozolomide and survival. Similarly, “Contrast Enhancement” and “Mass Effect” imaging features were found to be strongly correlated with expression of groups of genes involved in hypoxia and proliferation, respectively [6]. However, a systematic classification of tumor phenotype features in terms of their underlying cell biological processes on a genome-wide scale in GBM remains absent, although clinical applicability of these image features will depend on knowledge about how these features are driven by tumor biological processes that determine disease progression.

In this study, we present an Imaging-Genomics analysis to investigate the associations of a large set of biological processes and presurgical diagnostic MRI derived quantitative volumetric tumor phenotype features, such as Necrosis or Edema, focusing on the publicly available GBM dataset from The Cancer Genome Atlas (TCGA). These analyses were tied to molecular subtypes in GBM and prognostics. Image based volumetric features provide noninvasive tumor phenotype information complementary to genomic technologies and clinical information, potentially allowing advances in patient stratification and clinical decision-making.

To investigate which biological processes drive volumetric tumor phenotype features in GBM, we performed a pathway analysis based on gene expression profiles using a preranked Gene Set Enrichment Analysis [14]. We compared these results to molecular subtypes in GBM and evaluated the prognostic value.

The translation of quantitative imaging data into defined clinical settings requires knowledge of how volumetric tumor phenotype features are driven by biological processes that determine the outcome of a patient. This study presents an Imaging-Genomics analysis of presurgical diagnostic MRI derived volumetric features in GBM to evaluate if tumor phenotype features are associated with underlying tumor biology. We found different features to be enriched for different sets of biological processes. Molecular subtypes of GBM were difficult to be predicted by volumetric features. However, four out of five features showed significant prognostic value.

As correlations among our volumetric features were low to moderate in general, this suggests that quantifying each of those areas individually yields complementary information about the tumor phenotype beyond the Total Tumor Volume (TV). Interestingly, we found most of the biological processes to be anti-correlated to pathway expression. The most prevalent pathways were apoptosis, immune system, and signaling pathways, which were observed mainly for Necrosis (NE), Contrast Enhancement (CE), and Tumor Bulk (TB); features that were also significantly prognostic. As those pathways are known drivers of survival outcome [21‐23], this hence explains why NE, CE, and TB were found to be prognostic as well. Importantly, all of these features performed significantly better than TV, which highlights that quantification of individual imaging features should be preferred over calculating only the total tumor volume. Our finding that NE is anti-correlated with immune response and prognostic is in line with Gevaert et al. [24], who also correlated quantitative imaging features of GBM areas to molecular data and who found significant imaging associations to approximately 20 pathways. This analysis, however, differs from our analysis in that Gevaert et al. investigated a single slice of a tumor (in axial view), whereas we performed quantification using the 3D tumor volumes.

Edema (ED) was the only feature that was correlated with homeostasis, cell cycling, and proliferation pathways. Surprisingly, ED was not prognostic in our analysis, although cell cycling and proliferation are known to be involved in carcinogenesis [25]. However, using the publicly available MRI scoring scheme VASARI (https://​wiki.​nci.​nih.​gov/​display/​CIP/​VASARI), Gutman et al. [26] found ordinal ED assessment to be not prognostic as well. Interestingly, in a related study by Diehn et al. [6], binary assessment of ED resulted in significant survival predictions. Similarly, a recent study indicated that an ED volume cutoff of 85’000 mm³ is a significant prognostic factor using Kaplan-Meier analysis [8]; however, the rationale for this cutoff was not given. Prognostic performance of quantitative ED features could increase in cohorts of extended sample sizes, as ED has been reported to be a univariate predictor of survival in a large patient cohort previously [27]. Furthermore, our analysis suggests that CE and NE are prognostic. This is partially in line with Gutman et al. [26], according to whom CE is prognostic, but NE is not. The contradictions between our results and the studies by Diehn et al. and Gutman et al. could be due to the nature of the ED and NE assessments, which in our analysis were continuous, but binary and ordinal in Diehn et al. and Gutman et al., respectively. Likewise, our methodology could be compared to a study by Jamshidi et al. [28], but comparison remains challenging as their analysis focused on binary imaging traits on a relatively small cohort of patients and a subset of oncogenic pathways only.

Although ED was the only feature that was not prognostic in our analysis, it was the highest predictor of molecular subtypes in GBM instead. This is likely because ED was the only feature that expressed a different volumetric size distribution across molecular subtypes. Similar indications have been given by Gevaert et al. [24], who found three out of four features that correlated with molecular subtypes to be quantitative Edema features. In general, we found volumetric features to be only moderate predictors of subtypes suggesting that subtypes do not generally alter the size composition of tumor areas in GBM. Furthermore, we could not confirm that the proneural subtype has lower proportions of CE as suggested by Gutman et al. [26]. Poor predictability of Verhaak molecular subtypes by relative cerebral blood volume using T2-weighted MRI has been also described by Jain et al. [29].

In our analysis, ratios of volumetric features were not significantly prognostic or predictive of GBM subtypes. Generally, many more biological processes were significantly associated with the feature ratios, usually showing a trend towards a mix of pathways associated with the individual features that the ratios were composed of (e.g., NE/TV were enriched for signal transduction and biogenesis). While our study associated MRI volumetric features with biological processes, molecular subtypes, and survival outcome using genome-wide data and aimed at explaining the rationale for why MRI derived volumetric features are associated with survival on a pathway level, other studies have focused on revealing specific genetic variations between MRI features and survival [30‐34].

Our analysis was limited to a retrospective dataset. To establish volumetric biomarkers in clinical applications, prospective evaluation of our results will be required. Biological significance could be further validated by analyses of complementary molecular data such as mutational or epigenetic data. Such analyses could provide further insight into why separate quantification of distinct volumetric tumor phenotype features yield different biological and prognostic information. We acknowledge that the prognostic and predictive performances of the volumetric features in the TCGA-GBM dataset were moderate. Heterogeneity of GBM tumors [35] could be an explanation for this, which limits the definition of a single molecular subtype especially on the basis of single-needle biopsy [36, 37]. As imaging approaches target the entire visible tumor, we however expect in future studies that prognostic performances will drastically increase when Imaging-Genomic cohorts with even larger numbers of samples and standardized image processing become available for GBM.

While our study focused on volumetric phenotype features, alternative definitions of imaging phenotypes are available. This may, for example, include tumor location as this determines the extent of possible resection and hence is a prognostic factor in GBM [38]. In addition to such semantic phenotypes, agnostic phenotyping approaches such as radiomics could be added [39].

In conclusion, quantitative imaging biomarkers hold great potential, as, unlike traditional biopsies, medical imaging is noninvasive and captures the entire tumor volume. As we have shown, a relationship exists between individual volumetric phenotype features describing local, clinically relevant subareas of GBM tumors and global expression of genes. Knowledge about how these specific tumor areas are related to underlying biological cell processes may allow for advanced patient stratification and treatment decision on the basis of standard medical imaging, but efforts in optimization of prognostic and predictive performances need to continue.