A Systematic Review on the Current Radiogenomics Studies in Glioblastomas, P

HR A systematic review on the current radiogenomics studies in glioblastomas, p. 32-44 J VOLUME 4 | ISSUE 3

Review Neuro/Head and Neck Radiology A systematic review on the current radiogenomics studies in glioblastomas

Sotirios Bisdas1,2, Evangelia Ioannidou1,3, Felice D’Arco4 1Department of Neuroradiology, The National Hospital for Neurology and Neurosurgery, University College London NHS Foundation Trust, London, UK. 2Department of Brain Repair and Rehabilitation, Institute of Neurology, University College London, London, UK 3Medical School, University of Ioannina, Ioannina, Greece 4Neuroradiology Section, Department of Radiology, Great Ormond Street Hospital for Children, London, UK

Submission: 19/3/2019 | Acceptance: 26/08/2019

Abstract

Glioblastomas (GBM) have one of the poorest progno- rived from both morphologic and functional imaging ses of any cancer. Current cutting-edge research aims biomarkers (radiomics) in brain. Radiogenomics is at- to pave the way for new non-invasive methods of di- tempting to probe any correlation between radiological agnosing brain tumours through innovative imaging and histological features and hopefully assess the phys- techniques and genomic information from tumour sam- iological heterogeneity and genetic alterations paving ples. Over the past few years, various whole genome se- the way to a holistic approach of the tumour metabol- quencing analysis has identified biomarkers and thus ic, pathophysiological and structural fingerprint. This gradually changed the way of diagnosing brain tu- systematic review aims to summarise the current pub- mours. In this context, MRI is a versatile imaging tech- lished evidence of radiogenomics in GBM and also raise nique as it can provide multifaceted information de- awareness for future research in this field.

Key words Glioblastoma; Radiomics; Genomics; Biomarkers; Review; MR imaging

Corresponding Professor Sotirios Bisdas Author, Department of Neuroradiology, The National Hospital for Neurology and Guarantor Neurosurgery, Box 65, Queen Square 8-11, London WC1N 3BG, United Kingdom, Email: [email protected]

32 A systematic review on the current radiogenomics studies in glioblastomas, p. 32-44 HR VOLUME 4 | ISSUE 3 J

Introduction H3F3A gene, causing the encoding of histone H3.3; (v) A remarkable leap in the last decade has been the de- Phosphate and tensin homolog deleted on chromosome velopment of imaging techniques that help distinguish ten (PTEN) gene, also termed MMAC1 or TEP1, on chro- tumour from treatment effect, different tumoural mosomal band 10q23; (vi) Aberrant EGFR activity, re- grade and even different molecular profile. Radiom- sulting in EGFR overexpression; (vii) Receptor CX3CR1 ics is an emerging field that aims to extract quantita- and chemokine CX3CL1 positivity; (viii) O6-methylgua- tive data from medical images in order to characterise nine DNA methyltransferase (MGMT); and (ix) Vascular pathological processes [1]. Radiogenomics (aka imag- Endothelial Growth Factor (VEGF) overexpression. The ing genomics) in neuro-oncology uses radiomics to find role of the aformentioned molecular variations is two- correlation between images and molecular profile of fold. In addition to their use in categorising the large the tumour. Glioblastomas (GBM) manifest strong phe- variety of the glial tumours and contributing to a more notypic variations that can be assessed using magnetic holistic understanding of the pathophysiology and ma- resonance imaging (MRI), but still the majority of their lignant process, they are the foundation of “molecular- underlying biological drivers and genetic aberrations ly targeted therapy” and future of break-through imag- are largely unknown. Thus, efforts have been made ing techniques. over the past years to establish the role of radiogenom- Radiogenomics describes the correlation between ics in GBM as an important goal of this approach is the specific imaging phenotypes, using quantitative data ability to provide personalised therapy. This review and molecular characteristics of a certain disease. This aims to describe the current evidence for the added val- area is setting a new direction in oncology research. ue of radiogenomics in diagnosis and treating GBM and The question that subsequently emerges and we sought outline the premises of this emerging field in the future to address is to which extent are radiogenomics cur- of neuroradiology and neurooncology. rently elucidated. GBM are aggressive malignant primary tumours of the central nervous system (grade IV according to the Material and Methods WHO classification). They account for 45% of malignant Comprehensive, structured literature search was con- primary brain tumours [2]. Over 90% of diagnosed GBM ducted in PubMed for English articles published from cases are primary gliomas, arising from normal glial 2007 to 2018 on radiogenomics in oncology with search cells through multistep oncogenesis. The remaining terms including “radiogenomics”, “imaging genomics”, 10% are secondary gliomas originating from tumours “glioblastoma”, “genomics“, “gene” and “molecular”. of lower grade [3-6]. The aetiological background of Robust inclusion/exclusion criteria were applied for se- GBM has not been fully clarified, however the major- lection of eligible articles. Two authors separately per- ity of them are believed to be of spontaneous origin. formed quality assessment according to the QUADAS-2 The breakthrough in genetic identification in GBM was tool. Data were extracted in a pre-designed spread- achieved by Verhaak et al. [7], who distinguished four sheet following the PRISMA flowchart. References of different molecular subtypes in accordance with the included articles and literature reviews were checked genetic aberrations variability and gene-expression: for additional eligible studies. The principal aim was the classical, mesenchymal, proneural and neural sub- to include original and prevailing studies in humans type. The exact classification of GBMs is indeed very shedding light on the current position of radiogenom- challenging and clearly illustrates the need for new im- ics in the field of neurooncology and especially in GBM. aging surrogates of the molecular profile [8]. Although Studies appraising post-radiation-therapy imaging dis- extensive research has been ongoing for many years, tinctions as well as studies performing radiogenetics new GBM molecular biomarkers are discovered almost using immunochemistry were excluded, as this scope daily. These include: (i) Loss of 1p, 19q and 10q hete- of radiation genomics is clearly outside of our ongoing rozygosity; (ii) IDH1 or IDH2 mutations [9]; (iii) Elevated research and interest field. Furthermore, literature re- expression of epidermal growth factor (EGF), Iatrophi- ferring only to the term “radiomics”, without any el- lin, and "7-transmembrane domain-containing" pro- ements of “radiogenomics”, were not included in our tein 1 on chromosome 1 (ELTD1); (iv) Mutation in the review. In total, we included 23 articles, which con-

33 HR A systematic review on the current radiogenomics studies in glioblastomas, p. 32-44 J VOLUME 4 | ISSUE 3

tained original research into the current role of radiog- the images into standard stereotactic atlas space. enomics in GBM. The summary, including the unique The first comprehensive radiogenomic paper using value and any major shortcomings of the eligible stud- the open-access TCGA data base and TCIA images was ies are presented below. Illustrative examples of basic published by Zinn et al. [11]. Their analysis aimed to ex- radiogenomics correlation tasks for GBM genotypes plore genomic correlates of invasion and volume of the classification using multimodal MRI and the respective tumour. Peritumoural FLAIR and T2-weighted(w) im- textural features are shown in Figs. 1-4, where texture age signals were used to evaluate the extent of oedema images (intensities of the neighbouring voxels) are gen- and tumour infiltration, subgrouping the patients into erated using ITK-SNAP (http://itksnap.org). Compre- high, medium, and low FLAIR volume groups. High and hensive biomarker extraction can be performed using low groups were analysed and compared for differen- shape and intensity features, to capture tumour shape tial genomic expression profiles. The top upregulated or for distribution mapping through all voxels in the gene in the discovery (4-fold upregulation) and valida- segmented tumours. Haralick texture features relate tion (11-fold) sets was perostin (POSTN). The top down- to the tumour texture and include angular secondary regulated microRNA in both sets was miR-219, which moment, image contrast (large differences between binds and negatively regulates POSTN. Above median neighbouring voxels), entropy (the orderliness of the expression of POSTN resulted in significantly decreased gray level distribution in the image), correlation, sum survival and shorter time to disease progression. High square, sum average, inverse difference moment, sum POSTN and low miR-219 expression were significantly entropy, difference variance, sum variance, difference associated with the mesenchymal GBM subtype. How- entropy. ever, a limitation in the TCGA radiological data is the lack of image-tissue sample registration; thus, gene Results expression profiles cannot be matched to a specific The included studies and the investigated radiomics location on MRI [11]. Moreover, due to the large num- and genomics features along with any additional im- ber of probes, false positive gene hits may occur. This mune-histological biomarkers are illustratively sum- study and its outcomes were of great significance, as marised in Table 1. Ellingson et al. published an exten- they proposed a novel diagnostic method to screen for sive study on the association between GBM locations molecular cancer subtypes and genomic correlates of and imaging phenotypes, tumour molecular profiles cellular invasion. Last but not least, miRNA targeting and clinical variables in 507 patients with primary GBM is shaping the future of oncogenic alterations manip- [10]. The study demonstrates that the majority of GBM ulation. develops into the periventricular white matter regions A multilevel radiogenomic study by Jamshidi et al. adjacent to the subventricular zone. Moreover, results [12] detected GBM MRI radiogenomic signatures result- showed that MGMT promoter methylated tumours oc- ing from changes in messenger RNA (mRNA) expression cur regularly in the left temporal lobe, whilst tumours and DNA copy number variation (CNV). Imaging char- lacking loss of PTEN are found most frequently in the acteristics such as contrast enhancement, necrosis, frontal lobe. MGMT methylated tumours with the IDH1 contrast-to-necrosis ratio, T2 abnormality (infiltrative mutation tend to occur in the left frontal lobe. EGFR versus ooedematous tumour type), mass effect, and amplified and EGFR variant 3-expressing tumours occur subventricular zone (SZE) involvement were assessed. most frequently in the left temporal lobe. A similar re- A coherent MRI, mRNA and DNA radiogenomic associ- gion in the left temporal lobe was associated with ben- ation map for GBM was made, providing non-invasive eficial response to radiochemotherapy and increased evaluation of genomic signatures in patients with GBM. survival. Results from this study suggest that tumour For instance, the contrast enhancement association location may be related in the specific molecular and with LTBP1 could introduce enhancement as surrogate genetic profiling of GBM and their reaction to cytotox- biomarker for LTBP1 in more aggressive GBM pheno- ic treatment and overall survival. Nevertheless, critical types. The contrast-to-necrosis ratio was shown to be limitations to this study include: (a) the retrospective associated with RUNX3 and KLK, with increased ex- nature of the analysis and (b) the lack of registration of pression of RUNX3 expected to have inhibitory effects

34 A systematic review on the current radiogenomics studies in glioblastomas, p. 32-44 HR VOLUME 4 | ISSUE 3 J

a b

Fig. 1. T2-FLAIR (a) and textural features (b) images of an IDH-mutant, MGMT-methylated right temporal GBM.

Fig. 2. T2-w (a) and textural features (b) images of an IDH-mutant, MGMT-methylated and EGFR-amplified left temporal a b GBM. The post-contrast T1-w images (c) do not demonstrate any discernible enhancement. The post-contrast T1-weighted -texture features (d) are also demonstrated.

c d

35 HR A systematic review on the current radiogenomics studies in glioblastomas, p. 32-44 J VOLUME 4 | ISSUE 3

Fig. 3. Post-contrast T1-w images (left) and textural features (right) of an extensive left temporal IDH-wild-type, MGMT-unmethylated and EGFR-amplified GBM with necrotic areas and rather limited oedema.

a b c

d e f

Fig. 4. T2-FLAIR (a) and textural features (b) images of an IDH-mutant, MGMT-unmethylated, EGFR-amplified right frontal GBM showing cystic/necrotic regions and crossing the midline. The ADC (c) and relative cerebral blood volume (e) maps, along with their texture features maps (d and f, respectively) are provided.

36 A systematic review on the current radiogenomics studies in glioblastomas, p. 32-44 HR VOLUME 4 | ISSUE 3 J

Table 1. Summary table of the included radiogenomics in glioblastoma studies with the investigated radiomics-genomics features and immune-histological biomarkers.

Publication Imaging and Immuno-Histological Study (Reference) Year Biomarkers Genes MGMT promoter methylation, Ellingson et al. (10) 2013 IDH1 mutation status IDH1, EGFR, MGMT, PTEN POSTN, CXCL12*, COL1A1, Zinn et al. (11) 2011 microRNA expression COLA63, GRB10, SRPX2 SVZ involvement, KLK3, RUNX3, RAP2A, Jamshidi et al. (12) 2014 necrosis, FOXP1, PIK3KP1, LTBP1, mass effect CHI3L1, LM03 Oedema, necrosis, TP53, PTEN, EGFR, NF1, Gutman et al. (13) 2013 non-enhancement, IDH1, ERBB2, PDGFRA, minor axis RB1, TP53, PIK3CA

Ependymal. involvement, MYC oncogene, NDUFB3, Colen et al. (14) 2014 DWMT, NDUFB5, UQCRH, NDUFS4, enhancement across the midline COX17, COX5B, ATP5J ki67, Barajas et al. (15) 2012 necrosis, enhancement, rCBV EGFR, PDGFRA, PTEN, Hu et al. (16) 2017 Enhancement CDKN2A, RB1, TP53 nCBV, Cho et al. (17) 2018 MGMT promoter methylation IDH1, IDH2, MGMT, ATRX IRF9, MRPS6 , IFI44L, APBB3, Qian et al. (18) 2016 XRCC1, DYNLL2, MED12, TULP3, MGMT promoter methylation ERCC1, MGLL, ATF7 GAP43, Gevaert et al. (19) 2014 WWTR1, necrosis, enhancement IL4 rCBV peri-tumour, Liu et al. (20) 2017 ki-67 labelling index, IDH, TP53, EGFR mTOR Enhancement, tumour volume, Ellingson et al. (21) 2018 MGMT promoter methylation MGMT EGFR amplification, Kickingereder et al. (22) 2016 CDKN2A loss, PDGFRA, MDM4, CDK4, MGMT promoter methylation NF1

Hong et al. (26) 2017 nCBV, nADC IDH, ATRX

Zinn et al. (27) 2018 POSTN expression POSTN, NFKB

Demerath et al. (29) 2017 ki-67, EGFRvIII IDH1, OLIG1, OLIG2, BCAN

Radiomics including surface area, Smedley and Hsu (30) 2018 volume, sphericity, major-minor axis, compactness

on cell migration and invasion. This was the first study nancies. The SVZ involvement trait was associated with that associated KLK3 with GBM, although it has previ- genetic abnormalities such as RAP2A, which is involved ously been implicated as a gene existing in other malig- in in vitro migration and invasion of glioma cells [12].

37 HR A systematic review on the current radiogenomics studies in glioblastomas, p. 32-44 J VOLUME 4 | ISSUE 3

However, the major limitation of the above study was hanced (DSC) perfusion MRI for each tissue sample lo- the very small sample size. cation were differentiated to histopathologic traits (cell Gutman et al. [13] investigated molecular, clinical density, tumour score, proliferation, architectural dis- and presurgical MRI data in 75 patients with GBM. The ruption, hypoxia and microvascular hyperplasia). The importance of this study is the established significant tumour samples from the contrast enhancing regions correlation between contrast-enhanced tumours and had a higher tumour score, cellular density, prolifer- Verhaak´s gene expression classification. In particular, ation and architectural disruption than the non-en- the proneural subtype class was enriched with GBM hancing regions [15]. The apparent diffusion coefficient that displayed low levels (<5%) of contrast enhance- (ADC) was inversely correlated with the tumour score, ment. The mesenchymal subtype was shown to have re- the proliferation rate and the architectural disruption, markably lower rates of non-enhanced tissue compared whilst the fractional anisotropy (FA) was positively cor- to other tumour subtypes, suggesting distinct growth related with delicate microvasculature as well as archi- properties of proneural and mesenchymal GBM. EG- tectural disruption. Low measures of tumour ADC were FR-mutant GBM were found to be significantly larger, linked to more aggressive histological findings and the whereas TP53 mutant GBM were smaller in FLAIR im- relative cerebral blood volume (rCBV) had a positive ages than the wild type. The proportion of contrast-en- correlation with composite tumour score, proliferation hanced tumour and the length of the tumour major rate, total microvasculature, necrosis and tumour cell axis (based on FLAIR) were both significantly associat- number per high-power field. One possible limitation of ed with poor survival. Overall survival was shown not this study is any unavoidable mis-registration between to be associated with the percentage of necrosis, pro- biopsy area and MR images, due to the intraoperative portion of oedema, part of non-enhanced tumour, or brain shift. Nonetheless, this study sheds light into the length of the lesion’s minor axis. relation between histopathologic features of GBM and Colen et al. [14] analysed a total of 104 treatment-na- multiparametric MRI tumour findings. ive GBM patients with available gene expression results. Hu et al. [16] explored the possibility of the use of A group of neuroradiologists reported certain invasion multiparametric MRI and texture analysis to indicate type imaging features using the VASARI set. These fea- regional genetic diversity through MRI-enhancing and tures covered deep white matter tract (DWMT) involve- non-enhancing tumour fragments. The aim was to as- ment, ependymal enhancement, pial enhancement, en- sess the territorial intratumoural heterogeneity of ge- hancing tumour crossing the midline, non-enhancing netic profiles as reported through various stereotactic tumour crossing the midline, tumour extension into biopsies within a single tumour. A strong connection cortex, definition of the enhancing margin, definition between regional EGFR status and rCBV was detected. of the non-enhancing tumour margin, proportion of Cho et al. [17] deployed radiogenomics profiling to oedema, presence and absence of cysts and haemor- establish MRI-associated immune cell markers in GBM: rhage [14]. Subsequently, patients were divided on 258 patients with an initial diagnosis of GBM were re- the basis of presence or absence of the invasion MRI cruited. Fourteen immune cell markers were chosen phenotypes. The results showed that patients (Class for RNA-level analysis. Quantitative parameters from A) with a combination of deep white matter tracts and FLAIR, contrast-enhanced T1-w, DSC and DWI data- ependymal invasion had an important decline in sur- sets were used. CD68, CSF1R, CD33 and CD4 levels were vival. Analysis showed mitochondrial dysfunction to highly positively related with normalised rCBV values, be the top canonical pathway in the patients with an while CD3e and CD49d showed a significantly negative aggressive GBM phenotype gene expression signature correlation with ADC values. In addition, CD49d came [14]. The MYC oncogene was estimated to be the prime out to be an independent element for PFS of GBM pa- activation regulator in Class A tumours. tients. This radiogenomics outline unveils the correla- Barajas et al. [15] scrutinised the association of his- tion between immune cell markers, CBV and ADC values topathological features with imaging findings in 119 and highlights the role of ADC as prognostic biomarker. tissue specimens acquired from 51 GBM patients. Con- The differentiation between pseudoprogression (PsP) ventional, DWI and dynamic susceptibility contrast-en- and true tumour progression (TTP) in GBM is a chal-

38 A systematic review on the current radiogenomics studies in glioblastomas, p. 32-44 HR VOLUME 4 | ISSUE 3 J

lenging task and Qian et al. [18] sought for possible ra- pathway however the radiogenomics associations were diogenomic biomarkers related to PsP and TTP inves- different in enhancing and peri-enhancing area of the tigating clinical records, longitudinal imaging features examined GBM. The peri-tumoural rCBV had better and genomics. A series of morphological features were prognostic value than genomic biomarkers alone. extracted from the contrast-enhanced and necrotic Ellingson et al. [21] examined the correlation be- regions on the contrast-enhanced T1-weighted flu- tween postoperative residual enhancing tumour vol- id-attenuated inversion recovery (FLAIR). Thirty-three ume, genes expression and OS. Postsurgical, residual possible genes were chosen based on their connection enhancing disease was quantified and multivariate to the imaging features, reflecting their relation to PsP egression models were used to determine the influence and TTP. This study provided the first substantial evi- of clinical variables, O6-methylguanine-DNA methyl- dence that IRF9 and XRCC1 genes can serve as the po- transferase (MGMT) status, and residual tumour vol- tential biomarkers for the development of PsP and TTP. ume on OS [21]. Researchers came to the important Gevaert et al. [19] derived various quantitative im- conclusion that postoperative tumour volume is an aging features of GBM lesions to create radiogenomic anticipating aspect for OS, irrespective of the type of maps associating these features with molecular data therapy, age and MGMT promoter methylation status from 55 patients with GBM using the TCGA/TCIA da- with crucial negative impact on the survival of patients tabase. Eighteen image features were included in the with newly diagnosed GBM. study and were analysed in more detail for the three Kickingereder et al. [22] aimed to assess the link types of Regions of interest (ROIs). ROIs corresponding between multiparametric and multiregional imaging to enhancing necrotic portions of tumour and peritu- traits with major molecular features. Results however moural oedema were drawn, and quantitative image showed no evidence of tumour location preference for features were derived from these ROIs. Three enhance- any of the examined molecular parameters. Univariate ment features were significantly correlated with sur- imaging parameter associations were established for vival, 77 significant correlations were found between EGFR amplification and CDKN2A loss, with both demon- robust quantitative features and the VASARI feature strating increased rCBV and rCBF values [22]. The au- set, and seven image features were weakly correlated thors found correlations between MR characteristics with molecular subgroups [19]. The result was the cre- and molecular qualities, but the strength of the corre- ation of a radiogenomics map to link images to gene lations was not sufficient for utilisation of optimised modules and presents a promising complementary learning classification algorithms for prognosis of mo- strategy toward non-invasive management of GBM. lecular characteristics in patients with GBM. Liu et al. [20] sought to examine the correlation be- Hong et al. [23] examined the relationship between tween DSC MR perfusion parameters and genomic MRI modalities and important genomic profiles in biomarkers of GBM with regards to their prognostic GBM. Qualitative and quantitative imaging character- value. The mean and maximal rCBV ratio of both the istics such as volumetrics and histogram analysis from peri-enhancing tumour region and enhancing tumour, rCBV and ADC were assessed on the basis of T2-w and the Ki-67 labelling index, mammalian target of rapa- contrast-enhanced T1-w images. The imaging frame- mycin (mTOR) activation, epidermal growth factor re- work of variable genetic profile categories were - bal ceptor (EGFR) amplification, isocitrate dehydrogenase anced, and regression analysis were used for marking mutation and TP53 were collectively assessed [20]. A imaging-molecular correlations. The IDH mutation major correlation between maximum rCBV and mTOR subgroup had a larger T2-w volume and a higher vol- was detected, while the peri-tumoural rCBV showed ume ratio between T2-w and contrast-enhanced T1-w major connection to mTOR after correction for gender images than the IDH-wild type group. Higher mean and EGFR status. Finally, age and peri-tumoural rCBV ADC values were seen in IDH mutant tumours. Tumours were found to be the two strongest predictors of over- with ATRX-loss showed higher 5th percentile ADC than all survival. Overall and importantly, this study showed the IDH-wild type, no ATRX-loss counterparts. PFS was that haemodynamic abnormalities of GBM were asso- the longest in the IDH mutation group, followed by the ciated with genomics activation status of mTOR-EGFR IDH-wild type, ATRX-loss groups [23]. Apart from the

39 HR A systematic review on the current radiogenomics studies in glioblastomas, p. 32-44 J VOLUME 4 | ISSUE 3

retrospective design of this study, possible bias might Smedley and Hsu [26] utilised machine learning to be introduced by obtaining data from different MR outline gene expression phenotypes and morphology in scanners, incomplete genetic data for some of the pa- pre-operative MRI of GBM patients. An autoencoder was tients and the non-inclusion of genetic factors that are trained in 528 patient datasets, each of them with 12042 known to affect the prognosis of GBM patients such as gene expressions. The autoencoder's weights were used EGFR expression and PTEN. to initialise a supervised deep neural network. This di- Zinn et al. [24] attempted to demonstrate association rected application was trained in 109 patients with between gene expression variability status and MRI-ex- both genetic and MR information. Twenty morphologi- tracted radiomics. Radiogenomic forecast and affirma- cal image traits were extracted for every patient, from tion were completed using the Cancer Genome Atlas and contrast-enhancing and peri-tumoural oedema. Results Repository of Molecular Brain Neoplasia Data GBM pa- showed that the neural network had decreased errors tients and orthotopic xenografts. Tumour phenotypes in anticipating GBM phenotypes indicating that neural were disjointed and radiomic features were derived by networks may have the ability to identify a variety of the use of developed radiogenomic sequencing conduit. predictive radiogenomic relationships than pairwise Patients and animals were separated based on periostin or linear methods. A major limitation of this work was (POSTN) expression levels. Radiomics were applied to the sample size as well as the fact that the authors did anticipate POSTN expression levels in patient, mouse not assess whether pretraining with an autoencoder and interspecies. The brain-tissue focused normali- was better than other types of dimensionality reduc- sation and patient-specific normalisation are unique tion methods (i.e. principal component analysis or gene to this study, providing comparable cross-platform, module creation). cross-institution radiomic features [24]. POSTN expression levels were not correlated to any qualitative or Discussion volumetric MRI framework. Radiomic traits undoubt- Our review suggests that despite ongoing advances edly anticipated POSTN expression status in patients. in multimodal imaging technologies involving novel However, limitations in this study existed and were: (a) agents and powerful protocols or computer-aided tools, no prospective validation using spatially matched im- there is a gap between imaging information and the age-guided brain tumour biopsies in patients and ani- underlying molecular and genetic mechanisms of GMB. mal models and (b) the voxel size of the human MRI is This information gap has both clinical and socio-eco- larger compared to the xenograft MRI. nomic consequences since in many cases only imaging Demerath et al. [25] aimed to determine perfusion, information is available rendering decisions such as diffusion and chemical shift imaging (spectroscopy) the administration of expensive treatments hard and biomarkers in GBM and to associate them with genet- risking under- or overtreatment. Considering the large ically decided patterns of structural MRI. Some of the number of scanners and the huge volume of imaging results showed that rCBV in comparison to peri-tu- examinations taken each year, one can easily under- moural FLAIR hyperintensity was found to be higher in stand the enormous socioeconomic benefit for patients infiltration than in oedema. Moreover, axial diffusivity if molecular/genetic information could be inferred di- alongside peri-tumoural FLAIR hyperintensity was low- rectly from medical image analysis. This would require er in severe than in mild mass effect. Myo-inositol was medical imaging information to be correlated with positively correlated with Ki-67 in contrast-enhancing molecular/genetic characteristics of disease allowing a tumour. Therefore, alterations in rCBV and axial dif- more precise diagnosis, therapy planning and disease fusivity could be further examined as to what extend monitoring. they are connected to angiogenesis and activation of To address this challenge of transferring imaging proliferation genes. This pilot study suffered however data into prognostic information, the emerging field of of limited genetic information (only the IDH1-R132H radiogenomics proposes a data analysis framework for mutation state was available for all patients), retro- extraction and analysis of multidimensional multipar- spective design, small population and high variability ametric medical image features for optimising the di- in ROIs placement. agnosis and prognosis of disease. As such, radiogenom-

40 A systematic review on the current radiogenomics studies in glioblastomas, p. 32-44 HR VOLUME 4 | ISSUE 3 J

ics allows access to non-intuitive molecular or genetic or near-total resection, available second-line treatment information hidden in the images such as CT, PET and regimes and any possible salvage surgery. MRI that can be analysed with methods borrowed from We acknowledge that, with a wide array of multi- data science. Within radiogeneomics two branches of modal multiscale imaging features provided from the research are emerging. Firstly, correlation analysis on radiogenomics framework, the integration of this het- the imaging and genomic features can give insight on erogeneous high-dimensional information and its di- their association. Secondly, and higher-order statisti- mensionality reduction into statistically relevant sub- cal analysis (e.g. via tailored optimised learning algo- groups are one of the major challenges and the heart of rithms) can predict clinical outcomes based on a com- the analysis framework. Different data streams includ- bination of radio-genomic features by “learning” from ing quantitative imaging data and relative biomarkers the complementary information they give. from such as texture, shape, histograms and wavelets There are abundant advanced and emerging imaging have been represented in a unified framework and dif- biomarkers in brain utilising the full wealth of informa- ferences in scale and dimensionality have been sought tion that MRI can offer and the current evidence shows to be addressed. There are major benefits to studying that linked imaging-genomics maps offer non-invasive multiple layers of a system and on multiscale dimen- GBM molecular profiling by means of correlation- be sions using the various functional “omics” methods, but tween MRI biomarkers (derived from anatomical and to obtain maximum utility, the generated data need to functional images) and a rather wide range of genes be carefully integrated. To reduce “curse of dimension- expression. There have been some attempts to corre- ality” as machine learning models are in general prone late progression-free and overall survival with imaging to overfitting problems, radiomics and deep features, biomarkers or surrogates of genotypes but these works there is no reference standard for the continuously have been validated in small patient populations. The evolving machine-learning field and hence no solid major shortcoming in the existing literature is the lack evidence on synthesising these cutting-edge “-omics” of large-scale studies that use radiomics or radiog- methodologies has been generated. The different data enomics and therefore it is difficult to establish the syntheses included fragmented imaging and genomic role of radiogenomics in clinical patient management data analysis with subsequent correlation with sur- and to exemplify these exciting developments and op- vival rates. The method is based on the attempt to de- portunities in, and the promising future of, personal- tect all pairwise correlations in a few sets of features, ised cancer care. which remain significant, but one could argue that the It is spearheaded by various groups that the rela- link between imaging feature and genetic information tionship of quantitative imaging features with genetic was forced through biased selection or the link to sur- data, and the extrapolation of prognostic and predic- vival, and not that the genetic marker is a predictor of tive data from clinical imaging, will allow radiologists the imaging features. Another attempt was to use the to potentially diagnose and stratify patients for treat- molecular genetic data to extract information for pre- ment based upon imaging features alone. Furthermore, defined locations in the genome. These are then the serial monitoring of radiogenomics/radiomics-derived target features to be predicted from the imaging infor- biomarkers will better allow clinicians to monitor dis- mation. Each molecular marker is here considered in- ease recurrence and treatment response, while help- dependently, missing possible intermolecular correla- ing to tailor targeted-therapies to the ever-evolving tions. Moreover, the reported machine-learning-based tumour genome. These views have been at the level of performance measures are based on 10-fold cross-val- feasibility studies and there is currently not any Level idation rather than nested cross-validation, the latter 1-2 evidence of a clinically applicable value of radiog- being considered as less-biased and more powerful. In enomics in selecting or tailoring patient treatment. The future, deep learning that provides an unprecedented management plan for GBM still remains largely based leap in algorithmic performance by implicitly learning on the tumour location, patient’s age, available mul- relevant features directly from data should be part of ti-parametric MRI or PET to map more accurately the integrated radiogenomics projects. It is hence obvious tumour extent, neurosurgical estimation of complete that the research frameworks in radiogenomics are

41 HR A systematic review on the current radiogenomics studies in glioblastomas, p. 32-44 J VOLUME 4 | ISSUE 3

Fig. 5. Heatmap with dendrograms and hierarchical clustering of radiomic and genetic features. It was generated using the clus- tergram command in Matlab and a publicly available data from the Matlab library. The purpose is to illustrate how a combination of radiomic features (as columns) and genetic features (as rows) can be combined to give clusters of data with either positive outcome (in this graph red) or negative outcome (in this graph green). The classification of the different clusters is done via supervised machine learning.

definitely not trivial and require teamwork from stake- The existing work has eluded at integrating mul- holders of medical and engineering backgrounds to ti-parametric MRI radiomics with genome-wide meth- avoid repeating mistakes made in one field in another ylation supported by machine learning algorithms, field. as a combination of methods for improved GBM and Last but not least, the efficient integration of “omics” brain tumours prognosis. We encourage such com- data and platform outputs via the concurrent employ- bining of radiomic and genetic (i.e. radiogenomic) ment of bioinformatics should provide a deeper under- data with optimised learning algorithms to “mine” a standing of multilevel connections including gene-gene collection of features emerging from these complex and gene-environment interactions involved in disease datasets and via hierarchical clustering of the data development and evolution. The next decade will see (e.g. Fig. 5) seek patterns within them. The emerg- the translation of these tools and approaches into pre- ing patterns as outcomes from the machine learning dictive and preventive “systems medicine”. Discovery framework can be used for the purposes of explain- of new combined structural/metabolic imaging and ing different behaviours and determining novel bio- (epi)-genomics status biomarkers may contribute to a markers (exploratory analysis), stratifying different better understanding of pathological pathways that in GBMs and brain tumours in general (classification turn could be translated into diagnostic and therapy analysis) or to evaluate correlations between differ- monitoring clinical tools for application to personalised ent prognostic variables (regression analysis). All of medicine, and potentially to the elucidation of new drug these three aspects have not been explored enough to targets after group analysis of tumour subpopulations. date. Therefore, there is immense potential in these

42 A systematic review on the current radiogenomics studies in glioblastomas, p. 32-44 HR VOLUME 4 | ISSUE 3 J

frameworks to be a diagnostic and prognostic tool and utation and peculiarity of this topic the method is not its use needs to be discussed more and incorporated “prime-time” for clinical patient management, remain- within radiological journals. However, since the core ing essentially a research tool that necessitates exten- of radiogenomics is extensive data science, and spe- sive validation. Last but not least, many radiogenomics cifically using complicated computer simulations, it is studies have been of hypothesis-generating nature and important that this is user-friendly and achieved in a diligent verification in independent cohorts has been way that it showcases clinical usefulness. lacking. It may therefore be advantageous to investigate furthermore the effects of radiogenomics in the charac- Conclusion terisation and management of GBM as well as to point Radiogenomics, an amalgam of genomics and quantita- out the clinical validity and utility of any newly identi- tive medical imaging data, provide a continuously evolv- fied multiparametric MRI surrogate biomarkers. Radiog- ing unique, ground-breaking, non-invasive technique enomics have the power to expand radiology from a di- able to capture certain tumour characteristics that can agnostic and prognostic science to a discipline helping in help the non-invasive staging and give prognostic and elucidating new genetic pathways, establishing the value predictive information. The current research in radiog- of them in precision medicine. R enomics in GBM seeks mostly to establish diagnostic correlations between MRI-derived metrics and genes Conflict of interest expression profiles. Most of the findings reviewed here The authors declared no conflicts of interest. have demonstrated consistent beneficial outcomes from the entry of radiogenomics in neurooncology but there Acknowledgment: is still an inadequate number of studies in this area that The authors would like to thank Dr Jasmina Panovska-Griffiths for combine engineering/bioinformatics approaches and her valuable assistance and expertise in reviewing the manuscript radiogenomic datasets. Thus, despite the high-end rep- text and preparing the figures.

References

1. Kim Y, Cho H-H, Kim ST, et al. Radiomics features to Switz 2009; 19(3): 431–438. distinguish glioblastoma from primary central nerv- 6. Louis DN, Ohgaki H, Wiestler OD, et al. The 2007 WHO ous system lymphoma on multi-parametric MRI. classification of tumours of the central nervous sys- Neuroradiology 2018; 60(12): 1297-1305. tem. Acta Neuropathol (Berl) 2007; 114(2): 97–109. 2. Daniels LB, Shaya M, Nordberg ML, et al. Glioblas- 7. Verhaak RGW, Hoadley KA, Purdom E, et al. Inte- toma multiforme in two non-nuclear family mem- grated genomic analysis identifies clinically rele- bers. J La State Med Soc Off Organ La State Med Soc 2007; vant subtypes of glioblastoma characterised by ab- 159(4): 215–222. normalities in PDGFRA, IDH1, EGFR, and NF1. Cancer 3. Urbańska K, Sokołowska J, Szmidt M, et al. Glioblas- Cell 2010; 17(1): 98–110. toma multiforme-an overview. Contemp Oncol 2014; 8. Malik MT, Kazmi SJ, Turner S. Diagnostic challeng- 18(5): 307–312. es in primary brain stem glioblastoma multiform; a 4. Cancer Genome Atlas Research Network. Compre- case report. Interdiscip Neurosurg 2017; 10: 104–107. hensive genomic characterization defines human 9. Nicolaidis S. Biomarkers of glioblastoma multiforme. glioblastoma genes and core pathways. Nature 2008; Metabolism 2015; 64(3, Supplement 1): S22–27. 455(7216): 1061–1068. 10. Ellingson BM, Lai A, Harris RJ, et al. Probabilistic ra- 5. Rickert CH, Riemenschneider MJ, Schachenmayr W, diographic atlas of glioblastoma phenotypes. AJNR et al. Glioblastoma with adipocyte-like tumour cell Am J Neuroradiol 2013; 34(3): 533–540. differentiation-histological and molecular features 11. Zinn PO, Mahajan B, Majadan B, et al. Radiogenomic of a rare differentiation pattern. Brain Pathol Zurich mapping of oedema/cellular invasion MRI-pheno-

43 HR A systematic review on the current radiogenomics studies in glioblastomas, p. 32-44 J VOLUME 4 | ISSUE 3

types in glioblastoma multiforme. PloS One 2011; by using quantitative image features. Radiology 2014; 6(10): e25451. 273(1): 168–174. 12. Jamshidi N, Diehn M, Bredel M, et al. Illuminating 20. Liu X, Mangla R, Tian W, et al. The preliminary ra- radiogenomic characteristics of glioblastoma multi- diogenomics association between MR perfusion forme through integration of MR imaging, messen- imaging parameters and genomic biomarkers, and ger RNA expression, and DNA copy number varia- their predictive performance of overall survival in tion. Radiology 2014; 270(1): 1–2. patients with glioblastoma. J Neurooncol 2017; 135(3): 13. Gutman DA, Cooper LAD, Hwang SN, et al. MR im- 553–560. aging predictors of molecular profile and survival: 21. Ellingson BM, Abrey LE, Nelson SJ, et al. Validation multi-institutional study of the TCGA glioblastoma of postoperative residual contrast-enhancing tu- data set. Radiology 2013; 267(2): 560–569. mour volume as an independent prognostic factor 14. Colen RR, Vangel M, Wang J, et al. Imaging genom- for overall survival in newly diagnosed glioblasto- ic mapping of an invasive MRI phenotype predicts ma. Neuro-Oncol 2018; 20(9): 1240–1250. patient outcome and metabolic dysfunction: a TCGA 22. Kickingereder P, Bonekamp D, Nowosielski M, et glioma phenotype research group project. BMC Med al. Radiogenomics of glioblastoma: Machine learn- Genomics 2014; 7: 30. ing-based classification of molecular characteristics 15. Barajas RF, Phillips JJ, Parvataneni R, et al. Region- by using multiparametric and multiregional MR im- al variation in histopathologic features of tumour aging features. Radiology 2016; 281(3): 907–918. specimens from treatment-naive glioblastoma cor- 23. Hong EK, Choi SH, Shin DJ, et al. Radiogenomics correlates with anatomic and physiologic MR Imaging. relation between MR imaging features and major ge- Neuro-Oncol 2012; 14(7): 942–954. netic profiles in glioblastoma. Eur Radiol 2018; 28(10): 16. Hu LS, Ning S, Eschbacher JM, et al. Radiogenomics 4350–4361. to characterise regional genetic heterogeneity in 24. Zinn PO, Singh SK, Kotrotsou A, et al. A coclinical glioblastoma. Neuro-Oncol 2017; 19(1): 128–137. radiogenomic validation study: Conserved Magnet- 17. Cho HR, Jeon H, Park CK. Radiogenomics profiling ic Resonance radiomic appearance of periostin-ex- for glioblastoma-related immune cells reveals cd49d pressing glioblastoma in patients and xenograft expression correlation with MRI parameters and models. Clin Cancer Res 2018; 24(24): 6288-6299. prognosis. Sci Rep 2018; 8(1): 16022. 25. Demerath T, Simon-Gabriel CP, Kellner E, et al. Meso- 18. Qian X, Tan H, Zhang J, et al. Identification of bio- scopic imaging of glioblastomas: Are diffusion, perfu- markers for pseudo and true progression of GBM sion and spectroscopic measures influenced by the ra- based on radiogenomics study. Oncotarget 2016; diogenetic phenotype? Neuroradiol J 2017; 30(1): 36–47. 7(34): 55377–55394. 26. Smedley NF, Hsu W. Using deep neural networks for 19. Gevaert O, Mitchell LA, Achrol AS, et al. Glioblasto- radiogenomic analysis. Proc IEEE Int Symp Biomed Im- ma multiforme: exploratory radiogenomic analysis aging 2018; 2018: 1529–1533.

Ready - Made Bisdas S, Ioannidou E, D’Arco F. A systematic review on the current radiog- Citation enomics studies in glioblastomas. Hell J Radiol 2019; 4(3): 32-44.