Update on neuroimaging in brain tumours

Purpose of review To highlight some of the recent advances in magnetic resonance imaging (MRI), in terms of acquisition, analysis, and interpretation for primary diagnosis, treatment planning, and surveillance of patients with a brain tumour. Recent findings The rapidly emerging field of radiomics associates large numbers of imaging features with clinical characteristics. In the context of glioma, attempts are made to correlate such imaging features with the tumour genotype, using so-called radiogenomics. The T2-fluid attenuated inversion recovery (FLAIR) mismatch sign is an easy to apply imaging feature for identifying isocitrate dehydrogenase-mutant 1p/19q intact glioma with very high specificity. For treatment planning, resting state functional MRI (fMRI) may become as powerful as task-based fMRI. Functional ultrasound has shown the potential to identify functionally active cortex during surgery. For tumour response assessment automated techniques have been developed. Multiple new guidelines have become available, including those for adult and paediatric glioma and for leptomeningeal metastases, as well as on brain metastasis and perfusion imaging. Summary Neuroimaging plays a central role but still often falls short on essential questions. Advanced imaging acquisition and analysis techniques hold great promise for answering such questions, and are expected to change the role of neuroimaging for patient management substantially in the near future.


INTRODUCTION
Magnetic resonance imaging (MRI) is by far the most used imaging modality in patients with a brain tumour.Various distinct indications for brain tumour imaging can be identified, in particular primary diagnosis, focused treatment planning (surgery, radiation therapy), and surveillance, whether in the context of clinical routine or clinical trials.The present review highlights some of the recent advances in brain tumour MRI acquisition, postprocessing, analysis and interpretation for each of these indications.

Primary diagnosis of brain tumours: radiomics
Primary brain tumours are most commonly glioma.The classification of primary brain tumours and glioma in particular is a rapidly shifting landscape.In 2016, the World Health Organisation classification for central nervous system tumours for the first time includes molecular criteria alongside the traditional histopathological assessment [ Imaging research follows such developments closely.A fairly recently introduced but rapidly evolving field combines large numbers of (semi)quantitative imaging features with clinical characteristics in so-called radiomics research [3 The main adult glioma genotypes are those with and without isocitrate dehydrogenase (IDH) mutation (IDHmut respectively IDH wild type: IDHwt), those with IDH mutation further subdivided according their combined 1p and 19q status, such that three distinct tumour genotypes can be identified: IDHwt, IDHmut 1p/19q codeleted, and IDHmut 1p/19q intact glioma.Because each of these genotypes has clear prognostic and therapeutic characteristics, their accurate identification at the earliest possible stage, that is, even before tissue is available, is desirable and attempted with imaging.Imaging features of IDHmut 1p/19q codeleted tumours (traditionally known as oligodendroglioma) are well established, including calcifications, frontal lobe localisation, heterogeneity, cortical involvement, and indistinct margin.Such visual assessment on conventional imaging reaches about 50-79% accuracy [4 & ].A recently described and easy to apply finding to identify IDHmut 1p/19q intact glioma is the so-called T2-FLAIR mismatch sign (Fig. 1).Originally described in 2017 by Patel et al. [5], this sign has now been independently validated [6].A meta-analysis of five independent cohorts from four studies finds a posttest probability of a positive finding of 99% that IDH mutation is present, and 95% probability that the tumour is IDHmut 1p/19q intact [7 && ].Negative findings however only have a 74% probability that IDH mutation is absent and 29% probability that the tumour is not IDHmut 1p/19q intact.The sign is thus highly specific, albeit not perfect, but insensitive to identify IDHmut tumours and to differentiate these from 1p/ 19q codeleted tumours.An important implication of such imaging findings is, that even when tissue

KEY POINTS
Radiomics is a rapidly emerging field of imaging research, with the potential to aid primary diagnosis of brain tumour and their genotypes in the near future.
Advances in image analysis with artificial intelligence bring volumetric tumour assessment within reach and may allow for fully automated response assessment.
Perfusion magnetic resonance imaging is widely used for differentiating pseudo-from tumour progression; the newly published evidence-based recommendations for acquisition together with advanced analysis techniques are expected to improve both precision and accuracy.
Recent insights into the effects of repeated gadoliniumbased contrast agent (GBCA) administration result both in a re-evaluation of indications for contrast-enhanced scanning and in the development of image processing techniques that can substantially reduce or even omit GBCA administration.
Standardisation, facilitated by guidelines and recommendations, is key to building large imaging and metadata repositories, which in turn are crucial for the development and validation of AI solutions to further the field and role of neuroimaging.
diagnosis is known, these can still be used to determine patient management and predict patient outcome.Consider for instance the patient with IDHmut glioblastoma shown in Fig. 2. Based on traditional definitions gross total tumour resection has been achieved, since all contrast-enhancing tumour has been removed.Knowing however that-in contrast to IDHwt glioblastoma-IDHmut glioblastoma consist of a proportionally very large nonenhancing portion [8], it is clear that there is a substantial postresection tumour burden left, which needs to be taken into account for further treatment.
Automated algorithms, making use of large numbers of quantitative imaging features (radiomics) and deep learning techniques to predict tumour type and/ or grade are being developed and published in abundance [3 & ].An important limitation here, however, is the common lack of external validation and general scrutiny.Only 6% of over 500 studies published in 2018 of artificial intelligence (AI) algorithms was found to perform external validation [10].Evidently, the real-world performance of such algorithms can only be assessed by their application to previously unseen and entirely independent data.With the increasing availability of publicly available datasets as well as federated learning methods this could and should be readily achieved.
The holy grail of imaging is to provide diagnoses with such accuracy that tissue sampling solely for diagnostic purposes becomes obsolete.This is not yet within reach, but we do need to consider that the highest level of accuracy may not be required for specific patient groups.For instance, vulnerable patients in whom tumour biopsy is undesirable could already benefit from taking recent radiogenomics findings into account.In a population-based study, the clinical characteristics, care and outcome were studied in 131 patients with radiological diagnosis of glioblastoma without histological verification [11].As expected, these patients had extremely poor prognosis with median survival of 3.6 months, but those who received upfront temozolomide treatment did significantly better (with median survival of 6.8 months).This finding supports the notion that noninvasive diagnostic tools could be beneficial to guide treatment decisions even in patients too frail for tumour biopsy.

Treatment planning: functional imaging
Functional MRI (fMRI) remains an important adjunct to brain tumour surgery and intra-operative electrocortical mapping.Traditionally, task-based fMRI is used.This has several disadvantages, in particular its need for task performance, the fact that per scan of typically five minutes only one function or functional component can be assessed, and a complex set-up for stimulus presentation.Resting state fMRI (rs-fMRI) overcomes such issues, providing a more comprehensive view of functional networks in the brain, independent of task performance.The recommended scanning time is 10-12 min (in two runs to reduce head motion) [12].Several recent studies now indicate that-with particular acquisition and postprocessing techniquesrs-fMRI may perform as well as task-based fMRI in mapping the sensorimotor cortex and language system, with high sensitivity.A major disadvantage of rs-fMRI at present is its dependency on high levels of expertise, as analysis tools are typically still in the research domain and experience is not yet widespread.
Of particular interest is the emergence of supervised deep learning analysis techniques.In contrast to the more commonly employed independent component analysis, these analyses are entirely operator-independent as well as more suitable for individual subject's assessment, by using a-priori information on resting state networks of interest to assign a probability that each point in the brain belongs to such a predefined resting state network [12][13][14].In one study directly comparing taskbased fMRI and rs-fMRI of the language network using such an analysis, greater consistency across patients was found with rs-fMRI, a greater correlation with the language network, but also a greater degree of symmetry of this inherently asymmetrically represented brain function [15].This suggests that rs-fMRI may be less suitable for assessing language lateralisation, although the authors still found that there was a lateralisation in the posterior parietotemporal region and the contralateral cerebellum.Technically it is important to note that relatively slow acquisitions (repetition time [TR] of 1.5-3.5 s) may result in failure to reliably identify resting state networks in an individual subject, and high sampling rates are recommended (TR < 1 s) [16].Of course, rs-fMRI and task-based fMRI are not mutually exclusive, and by employing both techniques their respective strong points could be exploited.
Intra-operatively, researchers have successfully visualised functionally active cortex as well as tumour vascularisation using a technique coined mDoppler or high-frame rate Doppler ultrasound [17  & ].This technique has a spatial resolution of 50 mm, does not involve electrical stimulation greatly reducing the risk of intra-operative seizures, and seems to have a deeper penetration than conventionally used electrocortical stimulation mapping.It thus has great potential to improve intraoperative mapping of eloquent brain parenchyma and guide surgical resection as well as providing vascular information with high spatiotemporal resolution.

Surveillance: advances and guidelines
A hallmark paper on novel methods for response assessment, primarily in the setting of a clinical trial, is by Kickingereder et al. [18 && ] where they compared a fully automated assessment using AI to traditional manual response evaluation according to the Response Assessment for Neuro-Oncology (RANO) criteria.Validation of their results in an independent data set is under way.An important aspect of their work is the fully automated tumour segmentation.Numerous studies have indicated the benefits of volumetric assessment compared to conventional bidemensional measurements [19], but the lack of reliable, fast and user-friendly tools has thus far hampered their widespread implementation.Also, while adequate in the preoperative setting, currently available commercial tools for volumetric assessment still display considerable inter-rater variability due to the need for manual adjustment in more complicated settings, such as in the early postoperative phase and when assessing nonenhancing tumour burden [20,21].Kickingereder et al. 's publicly available algorithm HD-GLIO-AUTO (https://github.com/NeuroAI-HD/HD-GLIO-AUTO)provides an important step forward towards volumetric tumour surveillance, both in trials and in clinical practice.
A further important development is the ongoing standardisation of image acquisition and response assessment.The Response Assessment in Pediatric Neuro-Oncology working group published a series of three new guidelines: for low-grade glioma [22], high-grade glioma including diffuse midline glioma but excluding diffuse intrinsic pontine glioma (DIPG) [23 && ], and for DIPG [24].All three guidelines set minimum standards for image acquisition and specify response criteria.These guidelines clearly set themselves apart from the firmly established RANO criteria for adult glioma.One important aspect is the central role for nonenhancing lesions, acknowledging the fact that many paediatric glioma -whether high or low grade -do not show contrast enhancement, and even if they do, are often more reliably measurable on T2-weighted/T2-weighted-FLAIR images.A second aspect that is worth noting is the explicit incorporation of diffusion weighted imaging (DWI) in their response criteria for high grade glioma, albeit qualitatively [23 && ].In a way this is a revolutionary development, as this is the first deviation from solely relying on anatomical imaging for response assessment in patients with brain tumour.The guidelines furthermore provide guidance on real-world problems and dilemmas, such as the repeated use of gadolinium based contrast agent (GBCA) in view of concerns about its potential toxicity (see below), treatment related effects of novel therapeutic agents, distinguishing vasogenic oedema from infiltrative tumour, and measurement of cystic lesions.
In addition two guidelines were recently published on central nervous metastases.Kauffmann et al. [25] provide consensus recommendations for imaging brain metastases in clinical trials, outlining 'ideal' and 'standard' protocols.Le Rhun et al. [26] tested the score card for assessing leptomeningeal metastases (LMM) as recently proposed by the RANO-LMM group.They found that no acceptable agreement was found and that in the context of clinical trials, a centralised review therefore remains essential.The working group has since re-designed and re-evaluated the score card, results of which are expected to be published soon.

Gadolinium deposition
A further important aspect of tumour surveillance is the finding that repeated administration of GBCA is associated with deposition of gadolinium in the body [27].Although thus far there is no evidence of toxicity, there is a renewed awareness of the extensive use of GBCA.Also in light of the fact that there have been substantial improvements in MR image quality since indications for GBCA-administration were first defined, a re-assessment of the indications is in order.Two recent, retrospective studies with limited numbers of asymptomatic patients (N ¼ 29, N ¼ 18) under surveillance for untreated meningioma suggest that T2-weighted imaging provides similar measurements to postcontrast T1-weighted imaging for tumour follow-up and that GBCA-administration could thus potentially be omitted [28,29].
An exciting future perspective is to reduce [30] or even omit [31 && ] GBCA administration byusing AI to artificially construct 'virtual contrast-enhanced' images from multiparametric noncontrast MRI scans (Fig. 3).In a study of 47 enhancing tumours, 39 nonenhancing tumours, and 30 normal scans sensitivity and specificity of detecting enhancement on the artificially created virtual contrast-enhanced scans were 92% and 91% respectively, which is not perfect but still very impressive and holds incredible potential for reducing GBCA administration in the (near) future [31 && ].

Pseudoprogression: perfusion magnetic resonance imaging
An increase of radiological abnormalities that are indistinguishable from but do not constitute true tumour progression is called pseudoprogression and remains the most important diagnostic challenge in brain tumour surveillance [32].Pseudoprogression occurs after high-dose radiation therapy, with or without chemotherapy, and is now also being described after immunotherapy [33].This phenomenon highlights the fact that conventional imaging findings such as T2-weighted hyperintensity and contrast enhancement are nonspecific features of increased tissue water content and blood-brain barrier breakdown respectively.
Perfusion MRI is the most commonly applied advanced imaging technique to overcome these issues, based on the underlying pathophysiological principle that tumoral neovascularisation is associated with hyperperfusion, whereas inflammatory and other reactive changes due to treatment are not.Two meta-analyses report pooled sensitivities and specificities of 87-90% (95% confidence interval [CI], 0.82-0.94)and 86-88% (95% CI, 0.77-0.92) to distinguish pseudoprogression from the true progression of glioma, but it should be noted that almost all included studies are small, retrospective, and biased thus with overall low levels of evidence [34,35].Some important steps have been taken in assessing and improving the precision and accuracy of perfusion MRI.In the context of a multicentre clinical trial, precision was found to be low, indicating a need for standardisation of methodology and central review [36].A so-called digital reference object (DRO) was developed for evaluating the reproducibility of image acquisition and postprocessing software [37  & ].Applying this tool in twelve institutions from within the National Cancer's Institute's Quantitative Imaging Network, a substantial impact of both image acquisition and postprocessing was found on perfusion measurements, in particular in the context of a disrupted blood-brain barrier (i.e.enhancing lesions) [37 & ].Again using the DRO, it was found that using multiecho acquisitions were more robust than single-echo and obviate the need for a preload bolus which is commonly used to counteract the T1-effects of contrast-agent leakage into the extravascular space [38].Unfortunately, multiecho imaging is as yet not routinely available from the main scanner vendors.
A major step forward towards more uniform perfusion MRI is the recent publication of evidence-based consensus recommendations for dynamic contrast susceptibility imaging in highgrade glioma (Table 1) [39 & ].
From an analysis perspective, a newly described metric is fractional tumour burden.This not only removes some of the user dependency that is present with the current 'hot spot' technique, thus improving reproducibility, but also takes intra-lesional Neuroimaging heterogeneity into account reflecting the commonly present mixture of pseudo-and true progression [40,41].

Conclusion and future directions
Neuroimaging plays a central role but still often falls short on essential questions such as tumour type and grade, pseudo-versus true progression, and delineating tumours for focused therapy such as surgery or radiation therapy.Advanced imaging acquisition and analysis techniques hold great promise for answering such questions, but at present are too complex and expert-dependent for realworld clinical application.As these techniques become more mainstream, and smart approaches are applied to validate these, we can expect a big change in the role neuroimaging plays in patient management.Important drivers are the increasing awareness of the need for sharing of data, with the advent of large imaging and metadata repositories (e.g.https://www.cancerimagingarchive.net),and collaborative efforts such as the recently initiated European COST Action for glioma: GLiMR (https:// glimr.eu)[42].Finally, ongoing efforts to standardise terminology, acquisition, and postprocessing are not only important for clinical imaging in general, but will also result in high-quality, consistent imaging data to be aggregated for the development of AI technology.].BSW, Boxerman Schmainda Weiskoff as described in [43]; GRE-EPI, gradient recalled echo echo planar imaging; T1w þ C, postcontrast T1-weighted imaging.www.co-neurology.com

Update on neuroimaging in brain tumours
1].A further update of this classification is imminent with interim updates already incorporated in the recently published European Association of Neuro-Oncology guidelines on the diagnosis and treatment of adult glioma [2 & ].

FIGURE 2 .
FIGURE 2. (a) Pre-and (b) postresection contrast-enhanced Tl-weighted (T1w þ C) and T2-weighted (T2w) axial images of a patient with an IDHmut glioblastoma.The enhancing tumour portion (arrows) has been fully resected, but there is a large nonenhancing, T2weighted hyperintense (arrowheads) region, which should be considered residual tumour given the genetic status.
Weller M,van  den Bent M, Preusser M, et al.EANO guidelines on the diagnosis and treatment of diffuse gliomas of adulthood.Nat Rev Clin Oncol 2020; 18:170-1862020.Most recent classification scheme of adult glioma, including guidelines on diagnosis and management.3. && Lohmann P, Galldiks N, Kocher M, et al.Radiomics in neuro-oncology: basics, workflow, and applications.Methods 2020; 188:112-121.Comprehensive, excellent overview of radiomics analyses and review of radiomics studies in neuro-oncology.