Radiogenomic biomarkers for immunotherapy in glioblastoma: A systematic review of magnetic resonance imaging studies

Abstract Background Immunotherapy is an effective “precision medicine” treatment for several cancers. Imaging signatures of the underlying genome (radiogenomics) in glioblastoma patients may serve as preoperative biomarkers of the tumor-host immune apparatus. Validated biomarkers would have the potential to stratify patients during immunotherapy clinical trials, and if trials are beneficial, facilitate personalized neo-adjuvant treatment. The increased use of whole genome sequencing data, and the advances in bioinformatics and machine learning make such developments plausible. We performed a systematic review to determine the extent of development and validation of immune-related radiogenomic biomarkers for glioblastoma. Methods A systematic review was performed following PRISMA guidelines using the PubMed, Medline, and Embase databases. Qualitative analysis was performed by incorporating the QUADAS 2 tool and CLAIM checklist. PROSPERO registered: CRD42022340968. Extracted data were insufficiently homogenous to perform a meta-analysis. Results Nine studies, all retrospective, were included. Biomarkers extracted from magnetic resonance imaging volumes of interest included apparent diffusion coefficient values, relative cerebral blood volume values, and image-derived features. These biomarkers correlated with genomic markers from tumor cells or immune cells or with patient survival. The majority of studies had a high risk of bias and applicability concerns regarding the index test performed. Conclusions Radiogenomic immune biomarkers have the potential to provide early treatment options to patients with glioblastoma. Targeted immunotherapy, stratified by these biomarkers, has the potential to allow individualized neo-adjuvant precision treatment options in clinical trials. However, there are no prospective studies validating these biomarkers, and interpretation is limited due to study bias with little evidence of generalizability.

• Radiological biomarkers of key components of the tumor-host immune apparatus have been developed based on apparent diffusion coefficient values, cerebral blood volume values, or radiomics.
Radiogenomics focuses on the relationship between genomics and imaging phenotypes and is increasingly being applied in the research setting to characterize tumors which can be heterogeneous.2][3][4][5] Due to their infiltrative nature, diffuse gliomas typically have a very poor prognosis with the most common type glioblastoma, having a median Radiogenomic biomarkers for immunotherapy in glioblastoma: A systematic review of magnetic resonance imaging studies overall survival of only 14.6 months despite standardof-care treatment (which generally comprises surgery with maximal safe tumor resection, followed by radiotherapy with concomitant and adjuvant temozolomide chemotherapy). 6,7][10] Furthermore, in a randomized multicenter trial of recurrent glioblastoma, anti-programmed cell death protein-1 (PD-1) neoadjuvant immunotherapy has shown survival benefit. 11he challenge, however, is that the majority of patients in these studies have shown poor response to immunotherapy, attributable to the immunosuppressive tumor microenvironment (TME) with limited presence of immune cell populations.Current immunotherapies such as PD-1/PD-L1 inhibitors and chimeric antigen receptor T-cell therapy depend on the presence of these tumorinfiltrating lymphocytes within the TME, but these constitute only 10%-15% of all tumor-associated leukocytes. 12,13n addition, PD-1 expression in human glioma tissues is relatively low as compared to other cancers and is heterogeneous. 14Despite these challenges, there has been an increased interest in tumor-host immune apparatus target identification in glioblastoma. 9,11One such area of interest has been to identify preoperative imaging biomarkers that can stratify patients for neo-adjuvant treatment after diagnostic magnetic resonance imaging (MRI).
Early and noninvasive diagnosis and treatment therefore has the potential to improve patient quality of life and prolong survival.][3][4][5] Herein we systematically reviewed 9 studies that developed and validated MRI biomarkers that have the potential to be used, or have been used, for glioblastoma immunotherapy.The primary objective was to analyze immunerelated radiogenomic biomarkers.The secondary objective was to highlight alternative methods to develop immunotherapy biomarkers which were not radiogenomic.

Materials and Methods
We performed a systematic review (registered in PROSPERO; ID number CRD42022340968) of immunerelated radiogenomic biomarkers in glioblastoma.The search strategy followed Preferred Reporting Items for Systematic Reviews and Meta-Analysis (PRISMA) 15 (Figure 1; Supplementary Table S1).

Search Strategy and Selection Criteria
Search terms were applied to PubMed, MEDLINE, and EMBASE databases using medical subject headings (MeSH) terms 16 to identify original research articles published from January 1990 to January 2023 (Supplementary Table S2).A low-precision "high sensitivity search" 17 was conducted using subject headings and exploding terms.Studies not published in English, 18 editorials, conference proceedings, commentaries, letters, book chapters, laboratory-based or animal studies, preprints, or articles without peer review were excluded.

Inclusion Criteria
The patients studied were adults aged over 18 diagnosed with glioblastoma.All studies with abstracts where MRI was used to develop and/or validate biomarkers of the tumor-host immune apparatus were included.

Exclusion Criteria
All studies related to non-glial tumors; pediatric patients; vaccine trials; imaging other than MRI; and invasive studies including intratumoral injections or nanoparticle administration, were excluded.

Appraisal of Quality
The Quality Assessment of Diagnostic Accuracy Studies 2 (QUADAS 2) tool 19 was used to assess the quality of the studies focusing on risk of bias and concerns regarding applicability.Relevant items from the Checklist for Artificial Intelligence in Medical Imaging (CLAIM) were also used to appraise studies 20 (Supplementary Table S3).

Data Extraction
Data related to the type of study; MRI sequences; genomic markers; radiological markers, and their performance accuracy; and machine learning techniques employed, were extracted.Biomarkers were defined as diagnostic, prognostic, predictive, or monitoring according to the FDA-NIH BEST (Biomarkers, Endpoints, and other Tools) applied to neuro-oncology. 21

Importance of the Study
We present the first systematic review of immunerelated radiogenomic biomarker studies for glioblastoma.Radiological biomarkers of the tumor-host immune apparatus based on apparent diffusion coefficient values, cerebral blood volume values, and imagederived features including VASARI (Visually AcceSAble Rembrandt Images) and more complex radiomics have been developed within the last decade.The summarized evidence provides a basis to further develop and validate future immune-related radiogenomic biomarkers.If validated, these biomarkers have the potential to be further utilized for patient stratification during immunotherapy clinical trials for glioblastoma.

Results
Nine studies were included from 686 screened studies based on the PRISMA assessment (Figure 1).All studies [22][23][24][25][26][27][28][29][30] were retrospective and published after 2016 following the release of iRANO criteria for assessment of response to immunotherapy. 31Seven studies were radiogenomic and were the focus of the systematic review to achieve the primary objective (Table 1).The remaining 2 were non-radiogenomic (Table 2) but included for illustrative purposes to highlight how researchers can develop immunotherapy biomarkers without any association with genomic information (secondary objective).

Study Datasets
All studies included patients with histologically diagnosed "glioblastomas, isocitrate dehydrogenase (IDH)-wild type" or "astrocytoma, IDH-mutant, grade 4" who had undergone standard of care treatment. 6,34Among the radiogenomic studies, 6/7 (85.7%) were multicenter and one was performed using a dataset of 60 consecutive patients from a single center. 23The Cancer Imaging Archive (TCIA) MRI data (https://www.6][37] The study included 137 patients with TCIA MRIs, of which 46 had corresponding genomic information.In a second study, Jajamovich et al. 26 developed imaging biomarkers from 558 patients with TCGA genomic information, of which 50 had corresponding MRIs.In a third study, Liu et al. 27 used multiple datasets (TCGA, Chinese Glioma Genome Atlas, and Clinical Proteomic Tumor Analysis Consortium RNA-sequencing data; GSE13041 and GSE83300 RNA microarray data; TCIA and local institution imaging data) and developed biomarkers using a cohort of 774 patients with mRNA gene expression data from multicenter datasets including 70 patients matched with MRI and mRNA data (TCGA, Clinical Proteomic Tumor Analysis Consortium).Subsequently, the biomarkers were validated using MRI and survival data from a third independent cohort of 149 patients from a single center.In the fourth study, Rao et al. 28 studied 92 patients from the TCGA database with MRI, mRNA, miRNA, and survival data.In the fifth study, Narang et al. 29 developed biomarkers using matched MRI and mRNA data from 79 patients within the TCIA-TCGA database.The biomarkers were then trained on 35 patients and tested on 34 patients from a separate hospital cohort.Hsu et al. 30 identified biomarkers using matched MRI and mRNA data of 32 patients from TCIA-TCGA database and tested them on 84 patients with MRI Out of the 2 non-radiogenomic studies, one analyzed recurrent tumors 25 and the other included a mixture of newly diagnosed and recurrent tumors. 22Both studies included patients from immunotherapy clinical trials. 22,25Specifically, George et al. 22 used data from a multicenter phase II programmed death-ligand 1 clinical trial (NCT02336165) with a sample size of 113 patients partitioned into training and test sets.In the second study, Qin et al. 25 studied 10 consecutive patients enrolled in clinical trials of anti-PD-1 therapy with or without anti-CTLA-4 therapy (NCT02017717; NCT02054806).

Machine Learning, Radiomics, and Statistical Analysis
Eight studies (8/9; 88.9%) used manual or semi-automated segmentation for determining the image volume of interest and classified extracted image features with classical machine learning or advanced statistical modeling techniques while one study 28 did not use segmentation and applied VASARI (Visually AcceSAble Rembrandt Images) standardized features to advanced statistical modeling techniques.No deep-learning techniques were used.The extracted image features were either radiomicbased and obtained from structural images or consisted of quantitative ADC metrics.An exception was one study, which also extracted cerebral blood volume metrics in addition to ADC metrics. 23Radiomic features were extracted using Pyradiomics 24,27 (https://github.com/AIM-Harvard/pyradiomics) or the open source radiomics package by Vallières 22 (https://github.com/mvallieres/radiomics).
Liao et al. 24 used Pyradiomics to extract shape, first order, and texture-based radiomic features from 2D FLAIR images, and employed 4 different models on the data, namely Gradient Boosting Decision Tree (GBDT), logistic regression, support vector machine (SVM) and k-nearest neighbors (KNN).They showed that GBDT performance was best among the 4 models with an accuracy of 0.81 for classifying images into those related to short or long survivors.Six gene expression levels differed between the 2 survivor classes, 3 of which were moderately highly correlated with the most discriminative radiomic features.][37] Using a different approach, Jajamovich et al. 26 used MRI-derived ADC correlation analysis on gene expression data grouped into molecular subtypes as well as gene subgroups.The researchers demonstrated a negative correlation of mean ADC values with an immune-related gene signature subgroup containing CD4, CD86, and major histocompatibility complex class I and II which are associated with dendritic cell maturation, the complement system, and macrophage function.
Liu et al., 27 refined gene expression grouping further still using extracted shape, first order, wavelet, and texturebased radiomic features from intra-and peri-tumoral regions.Key features were selected using recursive feature elimination and SVM to generate a predictive model that classified tumors into those with low or high immune cell infiltration scores.These immune cell infiltration scores represented those immune cell infiltration patterns in the gene expression data that persisted in different datasets and were shown to be prognostic for survival.In an independent MRI dataset, the SVM model classified patients into predicted classes of low and high immune cell infiltration; only survival data was available as a reference standard.Rao et al. 28 used MRI VASARI features to dichotomize the data into 2 groups with corresponding scores according to the tumor volume class, T1/FLAIR ratio, and hemorrhage values.These radiomic groups were prognostic for survival and showed significant differences in gene expression levels within immune-related pathways (inducible co-stimulator (iCOS-iCOSL) signaling in T helper cells; retinoid X receptor (RXR) activation; and phosphoinositide 3-kinase (PI3K) signaling in B lymphocytes).

Downloaded
Narang et al. 29 obtained 6 radiomic-based imaging features (Gray-Level Size Zone Matrix, kurtosis, Neighborhood Gray Tone Difference Matrix) after feature selection tailored to gene expression levels of CD3 T cells using the Boruta algorithm.Using dichotomized CD3 counts, they trained and tested the classifier using the 6 features.A multivariate regression analysis demonstrated that the classifier was not confounded by clinical factors or tumor volume.
Hsu et al. 30 identified radiomic-based imaging features related to T1 CE and ADC images (first order, gray-level runlength matrix, gray-level co-occurrence matrix (GLCM)) that were able to classify clustering-derived immune cell subset patient groups based on immune profile combinations (cytotoxic T lymphocytes (CTLs), activated dendritic cells (aDCs), T regulatory cells (Tregs), myeloid-derived suppressor cells) using logistic regression models.The features were selected using random forest and information gain algorithms.

Radiological Imaging Biomarker Summary
Biomarkers extracted from MRI volumes of interest that correlated with various immune-related markers in patients with glioblastoma included ADC values, nCBV values, and image-based (VASARI, radiomics) features.
ADC biomarkers were negatively correlated with, firstly, CD3e and CD49d expression levels and, secondly, an immune-related gene signature (CD4, CD86, major histocompatibility complex class I and II) in respective studies. 23,26Similarly, nCBV biomarkers were positively correlated with expression levels of CD68, CSF1R, CD33 and CD4. 23Radiomic biomarkers (shape, first order, wavelet, and texture) were predictive of firstly, immune infiltration patterns/scores or CD3 expression levels in respective studies, 27,29 or secondly survival, which was shown to be correlated with immune-related genes (TIMP1, repressor of silencing 1, EREG), immune cell infiltration scores or other immune signatures, in respective studies. 24,27,30Simpler radiomic biomarkers (tumor volume-class, T1/FLAIR ratio, and hemorrhage phenotype) were predictive of survival, which was shown to be correlated with immune-related pathways (iCOS-iCOSL, RXR, and PI3K). 28Similarly, tumor volume was negatively correlated with CD123, CD49d, and CD117. 23The immune-related genomic and corresponding radiological biomarkers identified in this review are summarized in Table 4.

Bias Assessment and Applicability Concerns
A qualitative analysis of the risk of bias and concerns regarding applicability was performed for each study and is summarized in Supplementary Figure S1.Six (6/9; 67%) studies had a high risk of index test bias.The risk of bias was high or unclear in 6/9 (67%) studies regarding patient selection and was unclear in 4/9 (44%) studies regarding the reference standard used.Concerns of study applicability were high regarding the index test in 6/9 (67%) studies, high or unclear regarding patient selection in 7/9 (78%) studies, and unclear regarding the reference standard used in 5/9 (56%) studies.

Summary of Findings
The systematic review demonstrated that radiological biomarkers, namely ADC values, nCBV values, and radiomic features (VASARI, texture, shape, histogram, and wavelet) extracted from different MRI sequences, correlated with immune-related genetic markers and were developed as noninvasive radiogenomic biomarkers.Some studies used internal hold-out datasets for analytical biomarker validation 21 ; however, none used external hold-out datasets to validate the trained biomarker.Some non-radiogenomic biomarkers (ie, without any correlation with immunerelated genetic markers) were developed to predict response to immunotherapy.All reviewed studies are best considered as "proof of concept."

Studies Assessed
All studies employed retrospective designs.Limitations encompassed 6 main areas.
First, differences in the type of genomic data (single vs bulk RNA-sequencing data; microarray data, polymerase chain reaction or immunohistochemistry-based data) and their harmonization in each study confound pooled inferences from the different studies (Supplementary Table S4).
Second, patients underwent MRI imaging in different centers where there were differences in scanner manufacturer and local MRI sequence protocols.Different postprocessing steps were deployed in each study to tackle these differences but lack uniformity (Supplemental Table S5).It is plausible that there could be subsequent variability in the imaging features between centers confounding pooled inferences from the different studies.
Third, patient selection for the majority of studies was based on what had been included in public datasets (especially TCIA/TCGA) or small sets of local hospital data.Not only did the sample appear to be similar or the same in almost all studies (from TCIA/TCGA), but there was no clear and detailed explanation regarding the process of patient selection.For example, there was no clarity regarding patient selection being continuous or at random.Furthermore, other eligibility criteria varied amongst all the studies and again the details were unclear in the majority of studies.Confounded patient selection may mean that the study samples are not representative of the intended population ("glioblastomas, isocitrate dehydrogenase (IDH)-wild type" and "astrocytoma, IDH-mutant, grade 4") which limits the generalizability of the results

Neuro-Oncology Advances 13
Ghimire et al.: Radiogenomic biomarkers for immunotherapy in glioblastoma to the clinic.It is noteworthy that even if generalizable to the pooled grade 4 gliomas, the biomarkers developed in the studies have not been optimized for IDH-wild-type glioblastoma alone (as the datasets preceded the 2021 WHO classification).
Fourth, details regarding the reference standards used in the majority of studies were unclear and it would be challenging to reproduce them.Furthermore, tumor heterogeneity within the TME is likely to confound reference standards and may be a limitation in all the studies as the biopsy sample of the tumor, and subsequent tumortissue genomic data, may not entirely represent the overall TME of the tumor. 64,65The majority of the studies 23,24,[26][27][28] have not addressed other confounding variables such as age at diagnosis, resection status (biopsy, subtotal resection, total resection), postsurgical treatment (complete/ incomplete Stupp protocol) and second-line treatment including immunotherapy that are likely to influence the development and validation of prognostic biomarkers. 24,27,28oreover, diagnostic biomarkers can also be confounded by the unique interaction between the central nervous system, immune system, and advanced age in patients with glioma. 66An example relevant to 2 of the included studies 23,26 is that microglia express higher basal levels of MHCII and CD11b with age. 67ifth, the variable index tests developed as radiogenomic biomarkers did not undergo rigorous analytical validation and none were clinically validated. 21Internal hold-out test sets were used effectively to validate prognostic biomarkers after training in 2 studies 24,28 and a diagnostic biomarker in one study 29 (none were temporal hold-out test sets).Overall, these findings limit the generalizability of the results to the clinic.
Sixth, most studies employed indirect methods for biomarker development and validation.For example, an imaging biomarker might predict a gene expression signature; a separate dataset containing no imaging data might show that the same gene expression signature can predict survival.The separate dataset is not a hold test set for validating an imaging biomarker for either a gene expression signature or survival.The limitation is that such indirect methodology for imaging biomarker development shows there is some clinical relevance, but this is not analytical validation. 21Most studies likely employed such methods as there are few datasets containing imaging data that is matched with gene expression (for diagnostic biomarkers) or survival (for prognostic biomarkers).

Review Process
Pooled diffuse glioma (WHO grades 2-4) studies were excluded from the review process as it was beyond the research question, but we acknowledge that the biomarkers obtained in these studies might be of use in glioblastoma. 68,69ublication bias may have affected the range of performance accuracy of the biomarkers included in this systematic review.The potential for publication bias may be heightened by the omission of preprints and materials that have not undergone peer review.This is particularly relevant in the data science community, where the rapid pace of development often outstrips the slower process of peer review, leading some researchers to avoid submitting their work to peer-reviewed journals. 17The composition of the research team could therefore influence this bias.Teams with a stronger clinical focus might be more likely to seek publication in peer-reviewed journals, whereas those with a stronger emphasis on data science might not.

Study Explanations and Relevance From a National and International Perspective
The focus of most of these studies was on prognosis which may be of limited relevance to either identifying immune-related targets for immunotherapy; or for predicting therapeutic response to immunotherapy.1][72][73][74][75][76][77][78][79][80] Two areas of research can be combined to help develop panels of biomarkers which may be useful to stratify immunotherapy to treat particular tumors, and thereby contribute meaningfully to translation.[72][73][74][75][76][81][82][83][84][85] Second, there is an expanding arsenal of techniques to extract features including radiomics and deep learning features that can be used to develop imaging biomarkers in glioblastoma, [86][87][88][89][90][91] and even a decade ago non-immune radiogenomic glioblastoma studies demonstrated considerable promise. 92It is plausible that these 2 advancements, alongside an expanding number of new data repositories, may lead to the development of important biomarkers and allow translation to succeed-the review shows we are currently at a proof-of-concept stage.

Current Evidence in the Field
This is the first systematic review of immune-related radiogenomic biomarker studies for glioblastoma.One study that did not focus on glioblastoma patients but also included oligodendroglioma and astrocytoma patients, developed an immune TME radiomic signature. 93Here it was shown that the heterogeneity of the immune TME harbors prognostic impact.Other studies of interest have used different modalities.Nagle et al. 94 demonstrated imaging biomarkers for labeled CD8 T cells using positron emission tomography (PET) imaging in glioblastoma mouse models and showed the ability to quantify CD8 T cells noninvasively.Similarly, various radiomic signatures associated with CD8 T cells were identified in a systematic review by Ramlee et al. 95 related to various cancers including glioma (high and low-grade), gastrointestinal cancer, head and neck cancer, hepatobiliary cancer, lung cancer, breast cancer, and melanoma and their respective CD8 T-cellrelated radiomic signature obtained from imaging modalities such as PET, CT, and MRI.
Large high-quality multicenter studies are possible and should be the standard to aim for in neurooncology.In other oncology disciplines, this has been demonstrated.For example, Sun et al. 96 developed and validated CT-derived radiomic biomarkers related to tumor-infiltrating CD8 T cells in patients included in phase I trials of anti-programmed cell death protein-1 (PD-1) or anti-programmed cell death ligand 1 (PD-L1) monotherapy for solid malignant tumors.Similarly, Trebeschi et al. 97 developed CT-derived radiomic biomarkers for predicting response to immunotherapy in advanced melanoma and lung cancer patients.It is also noteworthy that platforms such as ImaGene (https://github.com/skr1/Imagene)have demonstrated the potential for reproducibility of radiogenomic analysis with initial feasibility experiments analyzing invasive breast carcinoma, and head and neck squamous cell carcinoma. 98

Implications for Future Research and Clinical Practice
The present review has revealed an absence of high-quality studies regarding immune-related radiogenomic markers in glioblastoma with concerns regarding bias and generalizability.Future large, multicenter, prospective studies using radiomic or deep learning methods are required for the development and validation of pertinent biomarkers.It is plausible that features extracted from images of modalities such as advanced MRI (including permeability, perfusion, diffusion, chemical exchange saturation transfer), MR spectroscopic imaging, and PET might provide additional information on tumor biology and microenvironment.Future studies could also develop and validate biomarkers for either IDH-wild-type glioblastoma alone which likely has a unique immune TME (biomarkers for postbiopsy settings at recurrence or during immunotherapy treatment), 99 or for lesions that are suspected to be glioblastoma (biomarkers for prebiopsy and neo-adjuvant settings which might include enhancing lower grade gliomas and other mimics).Candidate biomarkers need to be clinically validated in the setting of prospective studies.Whether a clinically validated biomarker demonstrates impact when used in conjunction with an intervention would require the biomarker to be integrated into immunotherapy clinical trials such as the CheckMate 143 study. 10Even if prospective biomarker studies are clinically validated soon, for example, to provide a panel of diagnostic biomarkers ready for patient stratification in downstream research, the scarce level 1 evidence for immunotherapy benefit currently means that biomarker studies demonstrating impact (ie, validated predictive biomarkers) when used in conjunction with an intervention, are unlikely to emerge soon.
Future studies might also use spatial transcriptomics or single-cell sequencing to better understand the role of immune cells in disease progression and lead to the discovery of new classes for radiogenomic analysis.Ultimately, there is the potential to produce noninvasive imaging biomarkers for neo-adjuvant immunotherapy stratification as part of personalized medicine within the next decade.

3
Ghimire et al.: Radiogenomic biomarkers for immunotherapy in glioblastomaData AnalysisPG, a neurosurgeon with 6 years of clinical and research experience performed the literature search, which was independently reviewed by TB, a neuroradiologist with 15 years of clinical and research experience.Any discrepancies were resolved after discussion.A meta-analysis could not be performed due to a lack of sufficient homogenous data from the systematic review and marked heterogeneity in the methodology of studies.

Table 2 .
Non-Radiogenomic Studies in the Review Downloaded from https://academic.oup.com/noa/article/6/1/vdae055/7641063 by guest on 29 April 2024 and survival data from the TCIA database; limited mRNA data were also available in the test set.

Table 4 .
Immune-Related Genomic Biomarkers With Corresponding Radiological Biomarkers Identified in the Review