Challenges and opportunities for advanced neuroimaging of glioblastoma

Abstract

Glioblastoma is the most aggressive of glial tumours in adults. On conventional magnetic resonance (MR) imaging, these tumours are observed as irregular enhancing lesions with areas of infiltrating tumour and cortical expansion. More advanced imaging techniques including diffusion-weighted MRI, perfusion-weighted MRI, MR spectroscopy and positron emission tomography (PET) imaging have found widespread application to diagnostic challenges in the setting of first diagnosis, treatment planning and follow-up. This review aims to educate readers with regard to the strengths and weaknesses of the clinical application of these imaging techniques. For example, this review shows that the (semi)quantitative analysis of the mentioned advanced imaging tools was found useful for assessing tumour aggressiveness and tumour extent, and aids in the differentiation of tumour progression from treatment-related effects. Although these techniques may aid in the diagnostic work-up and (post-)treatment phase of glioblastoma, so far no unequivocal imaging strategy is available. Furthermore, the use and further development of artificial intelligence (AI)-based tools could greatly enhance neuroradiological practice by automating labour-intensive tasks such as tumour measurements, and by providing additional diagnostic information such as prediction of tumour genotype. Nevertheless, due to the fact that advanced imaging and AI-diagnostics is not part of response assessment criteria, there is no harmonised guidance on their use, while at the same time the lack of standardisation severely hampers the definition of uniform guidelines.

Introduction

Glioblastoma (GBM) is the most common adult-type diffuse glioma, which arises from glial cells and concerns the most aggressive and malignant primary brain tumour with astrocyte-precursors, classified as Grade 4 by the World Health Organization (WHO) grading system. In spite of advances in treatment, the prognosis remains poor with a median survival of 14–16 months.^1,2 Following the recently published 2021 WHO Classification of Central Nervous System Tumours v. 5, GBM comprises only isocitrate dehydrogenase (IDH) wild-type tumours. All IDH-mutant diffuse astrocytic tumours are considered a single type called astrocytoma, IDH-mutant with WHO grades ranging from 2 to 4.³ While the separation of astrocytomas into IDH wild-type and -mutated tumours is an important advance in GBM classification, all of the literature which is covered within this review is based on the 2016 WHO classification of central nervous system tumours⁴ (or previous versions). Therefore, when the term GBM is utilised, this will not only include IDH-wildtype GBM; this will also comprise the new astrocytoma Grade 4 IDH-mutant entity.

Due to the continuous optimisation and development of imaging protocols, the role and value of neuroimaging in the diagnostic work-up and treatment evaluation of GBMs has increased over the last years. Advanced neuroimaging aids to non-invasively provide more certainty about the prognosis and response to therapy, which is beneficial for treatment decision-making and counselling of the patient. Therefore, the omnipresent role of advanced imaging in GBM is undisputed and further consolidation is driven by various major clinical and scientific societies (e.g. the European Society of Radiology, the Radiological Society of North America, the European Association of Nuclear Medicine, and the Society of Nuclear Medicine and Molecular Imaging).

The purpose of this review is to provide an educational overview of advanced neuroimaging techniques in GBM. More specifically, advanced MRI techniques, including diffusion-weighted imaging (DWI), diffusion tensor imaging (DTI), perfusion-weighted imaging (PWI) techniques (both with and without the use of contrast-agents), and magnetic resonance spectroscopy (MRS), will be discussed. In addition, Fluorine-18 fludeoxyglucose (¹⁸F-FDG-) and amino-acid tracer positron emission tomography (AA-PET) are discussed. We focus on the challenges and opportunities of applying these techniques in clinical practice. Also, some new developments within the field and the potential use of new radiotracers will be addressed. Furthermore, we discuss the potential added value of artificial intelligence (AI)-based tools. To conclude, we discuss possible solutions to avoid pitfalls in the study design, data acquisition, and data analyses for future clinical studies.

Conventional MRI of glioblastoma

Brain MRI with conventional imaging sequences plays a pivotal role in the diagnosis and follow-up of glioblastoma. Traditionally, the conventional sequences used in neuro-oncology include T ₁weighted (T ₁W, T ₂W, T ₂W fluid attenuation inversion recovery (FLAIR), and post-contrast T ₁W (T ₁W + c) sequences.⁵ Anatomical details of the brain and the neoplasm within can be accurately evaluated as well as peritumoral oedema and disruptions of the blood–brain barrier (BBB). Glioblastoma is characterised by a heterogeneous appearance on T ₁W and T ₂W, which can be explained by necrosis, haemorrhage, soft-tissue mass, and tumoural vasculature. The combination of an irregular enhancing lesion with areas of infiltrating tumour and cortical expansion is highly suggestive of GBM (Figure 1). However, the use of conventional MRI sequences alone limits the differentiation of GBM from other intracerebral mass lesions with cystic or necrotic components (e.g. other neuroglial tumours, brain metastasis, or brain abscess). In addition, conventional MRI is limited with regard to differentiating high from low-grade glioma.

Figure 1.

Conventional imaging of glioblastoma. Exemplary axial (a) T ₁ weighted images (with motion artefacts), (b) post-contrast T ₁ weighted images (with motion artefacts), (c) T ₂ weighted images and (d), FLAIR images. Contrast-enhancing lesion with non-enhancing components suggesting necrosis. The lesion is surrounded by T ₂W/FLAIR hyperintense signal representing tumour infiltration and oedema. The lesion itself is observed to exert mass effect. FLAIR, fluid attenuated inversion recovery.

In GBM, conventional MRI sequences provide some hallmark features which can aid to predict molecular markers. For example, small regions of enhancement, a larger non-enhancing tumour portion, well-defined tumour margins, and T ₁W hypointense areas with suppressed FLAIR signal within its necrotic components are predictive of IDH1-mutation (Figure 2A).^6,7 In addition, a large volume of T ₂W abnormality and a higher ratio of T ₂W to T ₁W + c tumour components were also found to be correlated with IDH-1 mutation (Figure 2B).⁸ Tumour location of IDH-1 mutation GBM in the frontal lobe has been reported most frequently by various groups.^9,10

Figure 2.

Conventional MRI of two patients, one with astrocytoma IDH mutant WHO Grade 4 and one with glioblastoma IDH wildtype. First row (a–d) shows an astrocytoma IDH mutant WHO Grade 4 located in the right frontal lobe with involvement of the rostrum corpus callosum. It is observed that only small regions of enhancing tumour are present on post-contrast T ₁ weighted images (b; see arrow head in the enlarged section). On T ₁ weighted images (a), very small hypointense areas which are also hypointense on FLAIR images (d) can be observed. Exemplary focus is encircled in the enlarged sections. These regions reflect necrotising/cystic regions. In addition, extensive T ₂W hyperintense regions can be observed on T ₂ weighted (c) and FLAIR images (d) surrounding the limited area of contrast-enhancing tumour. Second row (e–h) shows a glioblastoma IDH wildtype located in the left temporal lobe. It can be observed that a classical appearance is present with a relatively large ring-enhancing region on post-contrast T ₁ weighted images (f). This contrast-enhancing tumour is surrounded by a similar amount of T ₁W hyperintensity; see T ₂ weighted images (g) and FLAIR images (h). FLAIR, fluid attenuated inversion recovery; IDH, isocitrate dehydrogenase.

Next to IDH, the second most-reviewed gene is O⁶-Methylguanine-DNA methyltransferase (MGMT). MGMT methylation status is another important biomarker because high activity of MGMT (i.e. unmethylated MGMT) is known to result in a reduced efficacy of alkylating chemotherapeutic agents (e.g. temozolomide). In high-grade gliomas such as GBM, MGMT methylation is less common as compared to low-grade gliomas.^11,12 On conventional MRI, hypermethylated MGMT tumours tend to have mixed-nodular enhancement in lesions which are non-temporally located (Figure 3A).¹³ Unmethylated MGMT gliomas, on the other hand, show a ring-pattern enhancement (Figure 3B).¹⁴

Figure 3.

Axial post-contrast T₁ weighted images of patients suffering from glioblastoma with and without MGMT promoter methylation. (a) Axial post-contrast T ₁ weighted images of three patients suffering from glioblastoma with MGMT promoter methylation showing a predominantly mixed-nodular pattern of enhancement. (b) Axial post-contrast T ₁ weighted images of three patients suffering from glioblastoma without MGMT promoter methylation showing a predominantly ring-enhancement pattern. MGMT, O⁶-Methylguanine-DNA methyltransferase.

In non-GBM, two radiological signs on conventional MRI sequences have been described in literature, which can provide insights in the mutational status. First, the T2-FLAIR mismatch sign describes that areas with T ₂W-high signal intensity of the tumour are relatively hypointense in signal on T2-FLAIR images due to incomplete free water suppression. In addition, a rim of hyperintensity can be seen on FLAIR. These MRI features are considered a specific radiogenomic-signature of diffuse astrocytoma (IDH-mutant, 1p/19q intact) with a high positive predictive power.^15,16 The second radiogenomic-signature in non-GBM concerns the aspect of the T ₂W-hyperintense signal and its delineation from the normal brain parenchyma. When this hyperintense area has smooth borders and has a homogeneous signal intensity, the tumour is more likely to be an astrocytoma without 1p/19q co-deletion.^17–19

In the post-therapeutic setting, it has been advised to perform MRI within 2 days after surgical intervention to assess the extent of the resection, the presence of residual tumour, and the occurrence of post-surgical complications.²⁰ In the post-operative setting, blood products in the resection cavity may be mistaken for residual enhancing lesion due to its intrinsic T1 shortening effects. Therefore, pre- and post-contrast T ₁ weighted images must be evaluated with care. Nevertheless, enhancing lesions with a nodular aspect indicate residual neoplasm. Although these characteristics are generally reliable to assess residual or recurring neoplasm, some exceptions exist. Different treatment methods (i.e. chemotherapy and radiation) affect the permeability of the vascular walls which may lead to new enhancing lesions. If this contrast-enhancement is the result of treatment-induced vascular leakage, this is called pseudoprogression (PsP), whereas contrast-enhancement reflecting tumour recurrence is tumour progression (TP). Differentiation between PsP and TP is poor with conventional MR images alone (Figure 4). A 2011 study investigated the diagnostic accuracy of 11 signs as visible on conventional MRI to distinguish TP from PsP: 1) new enhancement; 2) marginal enhancement around the surgical cavity; 3) nodular enhancement; 4) callosal enhancement; 5) subependymal enhancement; 6) spreading wave front of enhancement; 7) cystic or necrotic change; 8) increased peritumoural T2 abnormality; 9) diffusion restriction; 10) decreasing enhancement intensity; and 11) increasing cystic or necrotic change. Only subependymal enhancement was found to have a limited predictive power with a sensitivity/specificity of 38/93% and with a negative-predictive value of 42%. The other 10 signs had no predictive power.²¹

Figure 4.

Conventional MRI of two patients with glioblastoma in the post-treatment setting. Axial post-contrast T ₁ weighted images of two patients with glioblastoma IDH wildtype. The upper images (a) show a region of contrast enhancement adjacent to the resection cavity in the right parietal lobe (white arrowheads). Radiological and clinical follow-up showed that this lesion represented pseudoprogression. The lower images (b) show nodular enhancement adjacent to the resection cavity in the left temporal lobe (white arrowheads). Radiological and clinical follow-up showed that this lesion represented tumour progression. IDH, isocitrate dehydrogenase.

Diffusion-weighted and diffusion-tensor MRI of glioblastoma

DWI is based on the random Brownian motion of water molecules and the magnitude of this Brownian motion is estimated as the apparent diffusion coefficient (ADC; mm²/s). In turn, ADC is dependent on the cellular density and the presence of macromolecules or organelles/cell membranes in tissue compartments.²² In gliomas, an inverse correlation between ADC values and tumour grade has been described.^23–25 DWI can also help to detect early tumour recurrence in enhancing and non-enhancing lesions seen as reduced diffusion²⁶ and to predict overall survival and progression-free survival in patients with glioblastoma.^27–30 DWI has been proposed to distinguish MGMT methylation status as a median ADC_min value of 800 × 10⁻⁶ mm²/s or higher was found to represent methylated MGMT status.³¹ DWI is, however, most commonly used to distinguish a brain abscess from glioma, and in the post-operative phase to identify areas of ischaemia.

In the post-therapeutic setting, differentiation between PsP and TP based on ADC values has been investigated by various groups (e.g. ^32–42). In general, it can be concluded that PsP lesions show a higher mean ADC-value as compared to the mean ADC-values in the TP group²⁶ (Figure 5). Although most studies on the use of ADC in the post-operative setting were conducted following a decent methodology, meta-analysis of ADC is hampered as different publications report different ADC metrics (e.g. mean, median, maximum, minimum). , When a quantitative assessment of ADC is carried out, it is recommended to use the mean ADC value of a region of interest (ROI), where necrotic areas should be excluded from the ROI-measurements. Reported cut-off values of mean ADC values to distinguish TP from PsP ranged between 1000 × 10⁻⁶ and 1412 × 10⁻⁶ mm²/s corresponding with a range in sensitivity and specificity of 78–98.3% and 63.6–100% respectively.^43–46 The highest accuracy reported in literature (sensitivity/specificity rate of 98.3/100%) was reported by using a cut-off value of mean ADC of 1313 × 10⁻⁶ mm²/s to differentiate between PsP and TP with higher values reflecting TP.⁴⁴ It must be emphasised that ADC values in post-treatment gliomas depend on a variety of factors, including post-operative artefacts (e.g. pneumocranium) and MRI system-related specifications such as magnetic field strength and b-values.³⁷ Therefore, it is recommended to perform ROI-analysis on fixed locations at different time points, as this allows one to assess the longitudinal changes in ADC values.

Figure 5.

Diffusion-weighted imaging in tumour recurrence in a patient with glioblastoma. Axial post-contrast T ₁ weighted images (a) show a new contrast-enhancing lesion cranial to the resection cavity in the right frontal lobe (white arrowheads). Restricted diffusion in this region is observed (white arrowheads) (b shows the corresponding ADC images; c shows the corresponding DWI images using a b-value of 1000 s/mm²). This was highly suggestive for tumour progression, which was confirmed by radiological and clinical follow-up. ADC, apparent diffusion coefficient; DWI, diffusion-weighted imaging.

DTI models complex tissue diffusivity, based on which the microstructural organisation of tissue can be evaluated. In DTI, additional gradient pulses are introduced which cause a random phase shift for diffusing molecules and cancel out stationary molecules.^47,48 In general, diffusion of water molecules in biological tissues tends to be anisotropic. The diffusion tensor can thus be represented by a diffusion ellipsoid with its main axis parallel to the principal diffusion direction within a voxel.^47,48 Different metrics can be derived from the DTI model (most used are mean diffusivity, MD and fractional anisotropy, FA). MD is comparable with ADC. FA serves as an index for the amount of diffusion anisotropy within the tissue; a value of 0 indicates isotropic water diffusion, whereas an FA value of 1 describes a maximally anisotropic voxel.⁴⁹ Some studies found that DTI metrics can be used to assess occult neoplastic invasion of white matter tracts^50,51 and the direction of tumour growth.⁵² In clinical practice, DTI is mainly used for tractography to guide neurosurgical procedures.

In the post-therapeutic setting, FA-values were found useful to differentiate PsP from TP in various papers with low or moderate risk of bias; these papers showed higher FA-values in TP as compared to PsP.^42,53–55 However, no prospective studies on this topic were found. Reported cut-off values of mean FA values to distinguish TP from PsP ranged between 0.13 and 0.18 and corresponded with sensitivity and specificity values of 68–81% and 73–79% respectively.^46,54,56 The highest reported sensitivity/specificity (81/79%) was found with a cut-off FA-value of 0.18, with lower values reflecting TP.⁵⁴ However, interpreting the FA-values suffers from similar limitations as ADC value interpretation.

Perfusion MRI of glioblastoma

The role of PWI of glioblastoma is mainly based on the presence of neovascularisation. Tumour neovascularisation concerns an extensive network of suboptimal, poorly organised vessels with slow flow and leakage.⁵⁷ On T ₁W + c images, areas of contrast-enhancement are seen. Slow flow and other dynamic capacities of the neo-angiogenic network can be assessed by PWI. The most commonly used PWI techniques are dynamic susceptibility contrast (DSC) perfusion, dynamic contrast enhancement (DCE) perfusion, and arterial spin labelling (ASL).

DSC PWI relies on the susceptibility induced signal loss on T ₂* weighted sequences, resulting from the passage of a bolus of gadolinium-based contrast agent. The most commonly used DSC perfusion parameter is cerebral blood volume (CBV) which can be estimated^58,59 and computed⁶⁰ based on the negative enhancement integral. Other parameters include cerebral blood flow (CBF), mean transit time (MTT) and time-to-peak (TTP). The estimated value of the area under the attenuation curve is proportional to the CBV but does not yield an absolute measurement. Therefore, the measurement is expressed relative to a standard reference, usually the contralateral white matter (relative CBV ratio: rCBV ratio).⁶¹ Overall, the rCBV ratio is an indicator of hypervascular regions and serves as the most robust parameter in DSC imaging.⁶²

DCE PWI relies on the evaluation of T1 shortening induced by a gadolinium-based contrast agent bolus leaking from the blood vessels in tissue. Pharmacokinetic modelling is used to derive various perfusion metrics including K^trans, V_e and V_p. K^trans represents the capillary permeability; V_e represents the fractional volume of the gadolinium-based contrast agent in the extravascular–extracellular space; V_p represents the fractional volume of the of the gadolinium-based contrast agent in the plasma space.⁵⁸

ASL is a perfusion technique without the need for contrast administration, where water molecules in blood vessels are tagged magnetically prior to entering the studied ROI (i.e. at the cervical level of the carotid artery). After a limited time interval (1.5–2.0 s), the labelled water molecules are imaged in the region of interest (i.e. the brain tissue). CBF values can be calculated from the differences in signal between the labelled images and the non-labelled images.^63,64 The major advantage of ASL is the fact that it does not suffer from contrast leakage effects.⁶⁵

In the diagnostic work-up, PWI is used for characterisation of glioma genotype, as it is known that genetic differences in glioma subtypes correlate with the glioma vasculature. An exemplary image of DSC PWI in the pre-operative setting is provided in Figure 6. A recent review and meta-analysis reported that DSC-derived CBV values were fairly accurate when predicting IDH genotype, with an area under the receiver operator curve (AUROC) of 0.83.⁶⁶ When reviewing DCE parameters an AUROC of 0.81, 0.84 and 0.78 were observed for K^trans, V_e and V_p, respectively. Insufficient data were available with regard to the non-invasive genotype prediction of GBM based on ASL perfusion metrics. However, ASL could be used to differentiate between tumour grade (Grade 2, 3, 4) according to one meta-analysis.⁶⁷ Furthermore, some studies suggested that pre-treatment rCBV_max values can be used as a prognostic marker for overall survival, or response to antiangiogenic treatment.^68–71

Figure 6.

Pre-operative MRI, including DSC perfusion weighted MRI, of glioblastoma. The upper row shows axial (a) T ₂ weighted, (b) T ₁ weighted and (c) post-contrast T ₁ weighted images of a patient with IDH wildtype glioblastoma. The lower row shows two axial perfusion maps derived from DSC perfusion-weighted MRI, namely (d) CBV and (f) CBF. The area of T ₁W-hyperintensity (presumably haemorrhage) shows low perfusion (white arrow heads), while the enhancing portion anterior to it shows increased perfusion (white arrows). CBF, cerebral blood flow; CBV, cerebral blood volume; DSC, dynamic susceptibility contrast; IDH, isocitrate dehydrogenase.

Most commonly, PWI is used in the post-therapeutic setting to aid in the differentiation between TP and PsP. A recent meta-analysis on this topic reported the diagnostic accuracy of two DSC parameters: mean rCBV and maximum rCBV. Pooled sensitivity and specificity for detecting TP were both 88% for a rCBV_mean ratio threshold ranging from 0.9 to 2.15. When using rCBV_max ratios, pooled sensitivity and specificity for detecting TP were 93 and 76%, with thresholds derived from literature ranging from 1.5 to 3.1.⁷² However, only the minority of papers included in these meta-analysis concerned prospective studies. An exemplary image of TP as assessed by DSC is shown in Figure 7.

Figure 7.

Tumour progression of glioblastoma in the left temporal lobe as appreciated on post-contrast T₁ weighted images and DSC perfusion-weighted MRI. Enhancing lesion in the left temporal lobe in a patient post-treatment which shows increased cerebral blood volume on DSC perfusion-weighted MRI (white arrowhead). These features are highly suggestive for tumour progression, which was confirmed by radiological and clinical follow-up. DSC, dynamic susceptibility contrast.

Regarding the use of DCE-PWI to differentiate TP from PsP, two recent meta-analyses reported a pooled sensitivity ranging from 89 to 92%. In both meta-analyses, the specificity was found to be 85%,.^72,73 However, only the minority of papers included in these meta-analyses concerned prospective studies. Additionally, pooled diagnostic accuracies were not based on one DCE parameter but considered overall reported diagnostic accuracy of DCE. Contrary to specific meta-analyses on DSC parameters, no meta-analysis has been performed on the diagnostic accuracy of either K^trans, V_e or V_p in DCE PWI post-therapeutic glioblastoma. Therefore, no range of thresholds can be provided.

The use of ASL to differentiate PsP from TP has been suggested to be less accurate in comparison to other PWI techniques according to a recent meta-analysis.⁷³ This meta-analysis reported a sensitivity ranging between 52 and 79% and a specificity ranging between 64 and 82% when ALS was used to differentiate PsP from TP. However, too few studies are available to perform a proper meta-analysis with pooled sensitivities and specificities and further investigation is warranted. However, a recent paper from our group reports that ASL and DSC have similar diagnostic accuracies suggesting that ASL could be an alternative for DSC-PWI. An example of ASL PWI in post-operative glioma is provided as Figure 8.

Figure 8.

Post-contrast T ₁ weighted images and ASL perfusion-weighted MRI overlay of glioblastoma after surgery. Panel a shows an axial post-contrast T ₁ weighted image illustrating contrast-enhancement of the resection cavity borders. Panel b shows the colour-coded ASL-derived cerebral blood flow overlay image, illustrating hyperperfusion in the contrast-enhancing parts. This indicated residual tumour/tumour progression. ASL, arterial spin labelling.

It has been reported that implementation of either DSC or DCE in routine follow-up MRI of GBM can aid the detection of tumour recurrence.⁷⁴ However, as stated above a wide range of cut-off values for each technique has been reported, which complicates further clinical implementation (e.g. ^75–77). Also, a variety of PWI metrics has been used in imaging trials using either PWI technique. A standardised perfusion scanning protocol and standardised methods data processing with validated criteria for the diagnostic work-up and follow-up of gliomas would contribute to more robust scientific and clinical data.⁶⁰ Efforts for standardisation of PWI acquisition have been made by various organisations (e.g. the American Society of Functional Neuroradiology)⁷⁸ and scientific papers.^60,79 Based on extensive simulations combined with expert knowledge, recommendations have been formulated with regard to DSC PWI (e.g. full-dose preload, full-dose bolus dosing using an intermediate (60°) flip-angle and choosing a field strength-dependent echo time (40–50 ms at 1.5 T, 20–35 ms at 3.0 T) in order to obtain overall best signal and precision for CBV estimates.⁶⁰ No consensus recommendations with regard to the use of DCE in neuro-oncological imaging are available. Although recommendations have been published for the use of ASL, it must be emphasised that these were not specifically designed for perfusion of neuro-oncological disease {⁷⁹ #232}. In summary, this paper recommended the use of pseudocontinuous labelling and background suppression. Also, a segmented three-dimensional readout without the use of vascular crushing gradients has been recommended {⁷⁹ #232}.

Finally, it has been shown that a well-established image-review process needs to be applied upfront to assess perfusion metrics because repeatability and reproducibility were found to be below 50 and 10% respectively in one multicentre study using DSC PWI.⁸⁰ Therefore, it has been recommended to measure rCBV in the context of clinical trials by two experienced readers. In case of disagreement, an adjudicator could be involved to provide the final perfusion measurement.⁸⁰ This is clearly time-consuming and thus impractical for clinical practice. Therefore, we recommend that PWI should be evaluated by an experienced reader and should always be interpreted together with other MRI sequences. Furthermore, radiological evolution of the area of interest over time and the clinical context of the patient should be taken into consideration when assessing PWI data.

MR spectroscopy of glioblastoma

The chemical composition of the tissue can be evaluated by MRS. MRS can therefore detect specific metabolites in defined regions of interest/voxels.^81,82 For GBM imaging, the most important metabolites include choline (Cho) and N-acetylaspartate (NAA) (Figure 9). In the clinical setting, MRS is often performed as a single voxel technique or as a slab comprising several voxels (i.e. multivoxel spectroscopy, chemical shift imaging). However, as not the entire lesion can be assessed in three dimensions at once, single voxel and multivoxel spectroscopy are known to suffer from sampling errors and outcomes can be confounded by the heterogeneous content of glioblastoma. Relatively new ¹H MRS sequences including 3D-echo planar spectroscopic imaging (3D EPSI) allow acquisition metabolic maps with an excellent coverage and spatial resolution. The images acquired from 3D EPSI protocols can be co-registered with anatomical images (T ₁ weighted images). MRS and 3D EPSI have been described as useful imaging protocols to assess GBM metabolism and differentiation of TP from PsP.^83–87 Tumour tissue has an MRS signature of increased Cho due to the increased cell density and total cell membrane. Reduced NAA is seen due to decreased neuronal content and decreased neuronal viability. Thus, elevation of Cho and decrease of NAA is suggestive for TP, although Cho can also be elevated in patients who receive immunotherapy.^88,89 In a meta-analysis of 55 studies, MRS showed to be superior to other parametric MRI sequences (conventional, ADC, DSC PWI, DCE PWI) when differentiating PsP from TP with a pooled sensitivity/specificity of 91/95%.⁷³ However, this meta-analysis included a heterogeneous collection of included papers. For example, studies on MRS with single voxel and multivoxel protocols were included and compared. Furthermore, this meta-analysis made no distinction between the diagnostic capacity of the different metabolite-ratios (included were MRS studies focusing on Cho/Cre, Lac/Cho, NAA/Cho and Cho/NAA). Reported cut-off values for Cho/Cre ranged between 1.07–2.50.^90–92 The cut-off values reported for Lac/Cho and Cho/NAA were 1.05 (34) and 1.71 respectively (48). Therefore, further research is needed, preferably with well-established study reading protocols and cut-off values.

Figure 9.

Single voxel MRS of glioblastoma at 1.5T. Brain MRI with axial post-contrast T ₁ weighted (panel a) and T2 (panel b) images of a patient with a small T2 hyperintense space occupying cortical/subcortical lesion in the right frontal lobe with some contrast enhancement. Panel c shows the MRS spectrum of the lesion. The corresponding spectrum (acquired with TE = 20 ms) can be indicative of glioblastoma tissue. The viable tumour tissue within the region of interest is recognised by high Cho signal and decreased NAA, the spectral peaks coming from lipid signal (lactate/lipid peaks are seen around 1 ppm on the spectrum) indicate the presence of necrotic tissue, even though this is not visible on structural MRI. The most prominent signals are labelled in the spectral pattern corresponding to voxel a. Cho, Choline; Cr1/Cr2, Creatine; Ins dd1, Myo-inositol; NAA, N-acetylaspartate; TE, echo time.

With regard to prognosis prediction, a recent MRS study showed higher Cho/NAA ratios in the post-operative peritumoral oedema zone in patients with early tumour recurrence in those areas. A higher Cho/NAA ratio in the peritumoral oedema zone was considered to be associated with poor prognosis.⁹³ However, this conclusion still needs to be corroborated by other research groups. 2-Hydroxyglutarate is an oncometabolite of IDH mutant glioma and it has recently been reported that MRS could be useful for the determination of IDH mutation status based on the detection of elevated levels of 2-hydroxyglutarate.^94–96 A meta-analysis on this topic found a pooled sensitivity of 84% and a pooled specificity of 97% with regard to predicting IDH mutation status in GBM.⁹⁷ Although promising, MRS is best carried out at 3T (or higher), requires expert knowledge and/or specific semi-automated computer-aided diagnosis software⁹⁸ and is not as widely available as other advanced MRI techniques.

¹⁸F-FDG-PET imaging of glioblastoma

The use of ¹⁸F-FDG-PET imaging in neuro-oncological diseases is limited by the poor tumour-to-background ratio as normal, healthy brain tissue also shows a very high physiological uptake of this tracer. Nevertheless, ¹⁸F-FDG-PET can provide some useful information in GBM in research settings. For example, the maximum standardised uptake value (SUV_max) of ¹⁸F-FDG-PET imaging could help to accurately determine the genotype of the GBM^99–102 and could help to predict patient prognosis.¹⁰⁰ In addition, differentiation between GBM and its main differential diagnoses (e.g. brain metastasis and primary central nervous system lymphoma (PCNSL)) by use of ¹⁸F-FDG-PET imaging showed encouraging results.^103–106 When using cut-off values of SUV_max≥15, SUV_max≥12, and SUV_max of ≥9.35, a sensitivity/specificity of 88/100%, 100/71.4%, and 100/78.3%, respectively, was reported with regard to discriminating PCNSL from GBM.^107–109

In the post-operative setting, GBM imaging by use of ¹⁸F-FDG PET has been investigated by various authors to distinguish PsP from TP.^110–118 From these studies, ¹⁸F-FDG PET showed a sensitivity ranging from 68 to 100% and a specificity ranging from 33 to 100%. Reported cut-off values of SUV tumour-to-brain ratios (TBR) vary from 0.75 to 2.64 with higher values reflecting TP.^111–119 However, a more recent meta-analysis also found a statistically significantly lower pooled specificity when imaging high-grade glioma as compared to low-grade glioma (82 vs 90%).¹¹⁹ Therefore, we conclude that FDG-PET imaging plays a limited role in post-treatment GBM imaging. Although mostly retrospective studies were carried out on this topic,^113–115 some well-performed prospective studies are also available.^117,118

AA-PET imaging of glioblastoma

In contrast to FDG-PET imaging, radio-labelled amino-acids have a high TBR due to the increased amino-acid metabolism in GBM cells due to cell proliferation and extracellular matrix production. Therefore, AA-PET can be helpful in the imaging of GBM in the pre-operative as well as post-therapeutic setting. S-11C-methyl)-L-methionine (11C-MET), O-(2-18Ffluoroethyl)-L-tyrosine (¹⁸F-FET), and 3,4-dihydroxy-6–18F-fluoro-L-phenylalanine (¹⁸F-FDOPA) are the most widely studied tracers. In the pre-operative setting, AA-PET imaging serves three main goals: primary diagnosis/differential diagnosis, glioma delineation, and treatment planning.

For ¹⁸F-FET PET imaging, it has been shown that abnormal focal ¹⁸F-FET uptake leads to a high sensitivity of high- grade glioma detection.^120–122 More specifically, a TBR_max<2.5 in ¹⁸F-FET excludes a high-grade tumour with high probability.¹²⁰ It has furthermore been found that dynamic ¹⁸F-FET data could be very useful for glioma grading.^123–125 Dunet et al reported that a negative slope of tumour FET time–activity was the best predictor of high-grade glioma.¹²⁵ Although several studies have been carried out with ¹¹C-MET and ¹⁸F-FDOPA, these radiotracers were less accurate as compared to ¹⁸F-FET.^126–128 In terms of glioma delineation, biopsy studies showed that ¹⁸F-FET PET-imaging detected the extent of gliomas, including GBM, most accurately.^129,130 With regard to glioma delineation, it is known from histological validation studies that conventional MRI is limited with regard to visualising glioma extent.^129,130 TBR_max values of ¹⁸F-FET PET imaging, on the other hand, were found to result in larger tumour volumes.¹³¹ Similar results have been obtained with ¹⁸F-FDOPA PET in progressive or recurrent GBM where a larger tumour extent was identified when compared with MRI-derived rCBV maps.¹³² Accordingly, ¹⁸F-FDOPA PET-based tumour volumes have been shown to extend beyond the contrast-enhancing volume on conventional MRI.^133,134. ¹¹C-MET has only been used in the delineation of recurrent GBM indicating contrast-enhanced MRI alone resulted in an underestimation of the tumour volume.¹³⁵

In the pre-operative setting, research showed that AA-PET imaging could be of predictive value as ¹⁸F-FDOPA was found capable to predict overall survival¹³⁶ and could identify bevacizumab-responders as early as 2 weeks after treatment initiation.¹³⁷ Concerning brain biopsy planning, the use of PET imaging to identify focal hot spots has also been found relevant for GBM imaging^133,138 (Figure 10), e.g. for detecting regions with abnormal activation of the EGFR gene due to a deletion of exons 2–7 of EGFR.¹³⁹ Also, AA-PET-CT was found to provide more accurate stereotactic biopsies compared to ¹⁸F-FDG-PET-CT guided biopsies.¹⁴⁰ Studies on the use of ¹⁸F-FET and ¹⁸F-FDOPA to guide stereotactic biopsies showed superior results compared to ¹¹C-MET and FDG-PET guided biopsies.^133,141,142 For instance, a recent cost-effectiveness analysis showed that the combined use of ¹⁸F-FET PET and MRI resulted in a 19% higher likelihood to obtain a representative biopsy.¹⁴² With regard to radiotherapy planning, the use of ¹¹C-MET PET imaging^143–145 and ¹⁸F-FET PET imaging effectively helped to delineate areas of glioma microspread resulting in a larger target volume.^131,146,147 Nevertheless, tumour recurrence after radiotherapy was reported to occur most often within the AA-PET defined target volume.^148,149

Figure 10.

¹⁸F-FET PET) images of glioma in the pre-operative setting¹⁸F-FET PET images show diffuse uptake of FET in the left frontoparietal lesion with focal areas of relative more uptake. These regions are thought to represent the localisations with the most high-grade tumour tissue. Biopsy was carried out and after histopathological examination, including molecular diagnostics, the diagnosis of an astrocytoma Grade III was made (according to the 2021 WHO Glioma Classification). ¹⁸F-FET, O-(2-18Ffluoroethyl)-L-tyrosine; PET, positron emission tomography.

Additionally, the aforementioned AA-based radiotracers are widely used in the post-treatment assessment, especially with regard to differentiation of PsP from TP. Several reports presented a sensitivity/specificity ranging from 66/60 to 78/100% when distinguishing PsP from TP by use of ¹¹C-MET PET imaging.^{90,119,150–159} When assessing the diagnostic accuracy of ¹⁸F-FET PET imaging (see, e.g. ^160–166), meta-analyses reported a pooled sensitivity and specificity of 88–91% and 78–95%, respectively.^119,167 The use of ¹⁸F-DOPA PET was to distinguish PsP from TP showed sensitivity and specificity ranges of 85–100% and 70–86%, respectively.^{119,168–171}

With regard to research on the use of AA-PET in imaging in post-operative glioblastoma, the field is limited by the large body of retrospective studies and the unblended assessment of radiologists/nuclear medicine physicians with regard to other clinical, histopathological and imaging information.^{90,113,157,158,164–166}

Novel MRI techniques for use in glioblastoma

During the last decades, tremendous developments of MRI hardware and image analysis methods allowed for micro- to macroscale imaging of GBM. These novel imaging approaches target a variety of molecular pathophysiological mechanisms occurring in GBM. To facilitate sharing of knowledge and to accelerate the clinical implementations of novel MRI techniques, the European, multidisciplinary network Glioma MR Imaging 2.0 (GliMR) was founded in 2019.¹⁷² Although the continuous development of novel MRI techniques prevents the authors from providing a complete overview of the literature on this topic, some promising techniques are highlighted.

Diffusion kurtosis imaging (DKI) has been proposed as a sensitive method to visualise the otherwise invisible, infiltrating component of GBM. DKI is an extension of DTI methods as it provides quantitative information about how tissue water diffusion deviates from a Gaussian distributed diffusion.¹⁷³ Previous research found that DKI variables can be used to assess microstructural alterations in perilesional white matter, suggestive of tumour infiltration which is not visible on conventional MRI sequences.^174,175

Another innovative imaging technique concerns amide proton transfer-chemical exchange saturation transfer (APT-CEST), which is aimed at assessing tumour metabolism and cell proliferation. Like other malignancies, GBM cells rewire their metabolism to grow excessively and to ensure prolonged cell-life. APT-CEST imaging indirectly visualises these mechanisms by detecting the presence of amide protons. Amide-protons accumulate in regions with an increased amount of proteins and peptides. APT-CEST imaging uses a specific radiofrequency pulse at the resonant frequency of protons inside amides (-NH). Thereby, only the protons within the amides will be saturated. In turn, the magnetic saturation of the amides will spontaneously be transferred to water due to the chemical exchange of the excited amide protons with non-excited protons within water molecules. The proton of the amides will thus be replaced with an unsaturated proton from water, causing an accumulation of saturation in water. Saturation of water will cause a decrease in water signal and is thereby an indirect reflection of amides in a target area. It has been reported that the use of APT-CEST is promising with regard to predicting IDH mutation status¹⁷⁶ and distinguishing TP from pseudoprogression.^177,178

Deuterium metabolic imaging (DMI) is novel, non-invasive approach which combines deuterium MRS with oral intake (or intravenous injection) of non-radioactive ²H-labelled substrates to generate three-dimensional metabolic maps. DMI can reveal glucose metabolism beyond uptake and thereby provides much more detailed information with regard to tissue metabolism as compared to ¹⁸F-FDG-PET imaging.¹⁷⁹ Preliminary data showed pronounced metabolic differences between normal brain and GBM.¹⁷⁹

Vessel architectural imaging (VAI) provides further insights into vessel size and type. VAI exploits the differences in observed proton relaxation from simultaneously acquired contrast-enhanced gradient recalled-echo and spin-echo MR imaging for vessel-size estimation. The temporal shift between the two relaxation curves can be used to estimate vessel type and size.^180–182 Because VAI can estimate these vessel features, it has been suggested that this technique might provide further insight into the mechanisms of pseudoprogression and the early detection of TP.^183,184

Novel radiotracers for use in glioblastoma

In recent years, several new classes of tracers have emerged that may also prove interesting for use in GBM not just from a diagnostic, but also from a therapeutic point of view as these tracers can potentially also be labelled with beta- and alpha-emitting radioisotopes. The first of these substance classes concern tracers aimed at the prostate-specific membrane antigen (PSMA). Primarily developed for use in diagnosis and therapy of metastatic prostate cancer, the inaptly named PSMA, which less colloquially is also known as glutamate carboxypeptidase two and in the brain serves as a modulator of excitatory neurotransmission, is also expressed in GBMs and their neovasculature in vitro as well as in vivo.^185–187 Although initial studies show promising results as to the diagnostic potential of PSMA in GBM,¹⁸⁸ no prospective data on the diagnostic value of this tracer are available as yet. However, in contrast to AA-PET, given sufficiently high uptake PSMA-PET as stated above also may indicate to possibility of radionuclide therapy with PSMA targeted substances.

The second class of interest in GBM are tracers aimed at the fibroblast activating protein (FAP). Recently introduced as a novel tracer of interest for diagnosis and possibly also therapy in a broad variety of oncological diseases,¹⁸⁹ this class of tracer in first, preliminary results also was shown to produce promising results in IDH-wildtype GBM as well as IDH-mutant astrocytoma, but not in low-grade IDH-mutant gliomas. This indicates that this class of substance may in the future play a role in non-invasively identifying high-grade IDH-mutant gliomas and GBMs, as well as may provide another radionuclide therapy option for these diseases.

Artificial intelligence and imaging of glioblastoma

A growing number of AI-based applications is finding its way to clinical practice. Commercially available AI-based software in the field of radiology holds at least 100 CE-marked products, although the majority of these packages have not been based on peer-reviewed scientific evidence and only a minority of such software applications have demonstrated (potential) clinical impact.¹⁹⁰ Although AI holds the potential to perform image analyses which matches and potentially surpasses the experienced neuroradiologist’s abilities,¹⁹¹ important challenges for AI need to be recognised (for a review see¹⁹²). For example, most studies on AI in GBM imaging were performed on limited sample sizes. Additionally, in most studies there was no external validation of obtained results, which limits generalisation of the study results and bears the risk of overfitting. It is therefore essential to educate the end-users on this topic, who should be familiar with the strengths and limitations for applying AI-tools in clinical practice. Also, data sharing is an important factor which will contribute to further development and implementation of AI-tools. For that reason, the open-access imaging databases of the MICCAI¹⁹³ and the Erasmus Glioma Database¹⁹⁴ are much needed advances in this field. Creating and maintaining such large databases are essential, though time-intensive tasks. AI could also play a role in data curation in order to preserve its integrity and ensure reusability.¹⁹⁵

The potential of AI tools in glioma image analysis is also immense. First, AI-based glioma segmentation was found to be highly accurate,¹⁹⁶ which would relieve radiologists of the labour-intensive task of image segmentation or even tumour measurement. This explains why fully automated segmentation tools have gained interest over the years and highlights the importance of the annual multimodal Brain Tumour Segmentation (BraTS) challenge (http://braintumorsegmentation.org/). A recent report showed that the implementation of an AI- based segmentation tool in our clinical practice resulted in reasonable (77%) rates of successful segmentation.¹⁹⁷

In one study, researchers used an AI application as an add-on feature in radiological readings to predict IDH mutation status in gliomas.¹⁹⁸ In this study, neuroradiologists’ predictive capacity of WHO grade was improved when coupled with the predictive capacities of a random forest algorithm. In addition, the same input variables could be used to accurately predicting IDH mutation status.¹⁹⁸ Two recent meta-analysis on this subject provided an overview of the accuracy of AI with regard to predicting glioma genomics, which showed high accuracy of machine learning algorithms for the prediction of IDH mutation status, 1p/19q codeletion status, MGMT promoter mutation and TERT promoter mutation with AUROCs of 0.909, 0.748, 0.866 and 0.842, respectively.¹⁹⁹ However, again these studies mainly used internal validation and external validation was largely lacking. In a recent review, similar predictive capacities were found with pooled sensitivities and specificities of 0.88, 0.76 and 0.76 for predicting IDH mutation status, 1p/19q codeletion status and MGMT promoter mutation, respectively.²⁰⁰ Pooled specificity for predicting IDH mutation status, 1p/19q codeletion status and MGMT promoter mutation showed to 0.86, 0.83 and 0.83, respectively. However, these meta-data represent the predictive capacities of the investigated artificial intelligence approaches in their training phase. With regard to the IDH mutation status, a meta-analysis was provided on the diagnostic accuracy parameters in the (mostly internal) validation sets. The pooled sensitivity and specificity for predicting IDH mutation status were 0.83 and 0.85 in validation sets, respectively.²⁰⁰

Recommendations and conclusions

From the published literature, it is evident that advanced neuroimaging techniques have the potential to add value to the diagnostic work-up, treatment planning and surveillance of patients with suspected GBM. It is, however, also clear that there is wide variability in their application and interpretation, while patient access to these techniques also varies widely.²⁰¹ With the exception of the newly defined response assessment criteria for paediatric low-grade glioma—which include DWI—advanced imaging techniques are not part of international and widely accepted response assessment criteria such as those defined by the Response Assessment for Neuro Oncology (RANO) working groups. This can at least in part be attributed to the lack of standardisation and high-level evidence. Most published work concerns retrospective, selected and often small patient populations, while prospective studies—in particular using external validation cohorts—are rare and randomised controlled trials non-existent. Here, we find ourselves somewhat in a Catch-22: due to the fact that advanced imaging is not part of response assessment criteria, there is no harmonised guidance on its use, while at the same time, the lack of standardisation severely hampers the definition of uniform guidelines. Recent international efforts to standardise image acquisition in glioma^5,202 are instrumental to break this vicious circle and to facilitate studies for obtaining the much needed high-level evidence for providing clear guidance on the optimal choice and application of the various advanced imaging techniques.

At the same time, it seems that PWI and MRS, as well as AA-PET have found widespread application for diagnostic challenges, in particular for assessing tumour aggressiveness, extent, and differentiation from treatment-related effects, while DWI has its particular use for differential diagnosis and pre-surgical tractography. Novel tracers aimed at PSMA or FAP could mix up the landscape here as these tracers also seem to have specific diagnostic properties as well as the potential for radionuclide therapy. It remains to be seen if and which technique is superior to others, and until such time individual institutions apply techniques according to local preferences and expertise. In the absence of a clearly superior technique, the local level of expertise is probably a more important factor than the choice of technique. This also means that the acquisition, post-processing, and interpretation need to be in the hands of experts who are not only technically but also clinically informed on the strengths and weaknesses of each applied technique.

The same could be said about AI-based tools, which have enormous potential to improve and enhance clinical practice, on the one hand automating tasks that are currently done manually (such as tumour segmentation) and on the other hand providing additional diagnostic information (such as prediction of tumour genotype). Again, levels of evidence are still low and the time is now ripe for studies with prospective, sufficiently powered cohorts and—crucially—external validation. Guidance for setting up and assessing the strength of AI studies is for instance provided by the radiomics quality score.²⁰³

The classification of diffuse glioma, including GBM, is a rapidly changing the landscape with an important role for diagnostic imaging at every step along the way from diagnosis to treatment decision-making and treatment monitoring. Advanced imaging acquisition and AI-based analyses provide powerful tools, both in their current form and in future developments, but need to be used with expert knowledge within the context of the currently available evidence.