The diagnostic role of diffusional kurtosis imaging in glioma grading and differentiation of gliomas from other intra-axial brain tumours: a systematic review with critical appraisal and meta-analysis

Gliomas are the commonest primary brain tumour and they remain a leading cause of solid cancer-related deaths in the under 40s [1]. Gliomas encompass a heterogeneous broad group of tumours with different cellular origins and variable biological behaviour. Classification of gliomas is therefore essential to guide therapy, anticipate treatment response and predict prognosis. Historically, the old WHO glioma classification was based on a histologic definition of predominant cellular lineage and grade, essentially differentiating gliomas into high-grade more aggressive and low-grade less aggressive [2]. The advent of biomolecular characterisation has led to the identification of key genetic markers which also influence tumour behaviour. The isocitrate dehydrogenase (IDH) gene has received particular recognition and has contributed to a major revision in the latest 2016 World Health Organization classification [3]. Despite these advances, this essential classification remains reliant on invasive tissue sampling. This has considerable risks, including foremost permanent neurological deficit which also greatly precludes repeat sampling to check for high-grade transformation. Tissue sampling is additionally fallible, with biopsies and incomplete resections potentially providing a non-representative sample due to intra-tumoral heterogeneity. Thus, there is a clear need for a non-invasive imaging marker of tumour type, grade and genetic status which would inform management and estimate prognosis.

Diffusion-weighted imaging (DWI) and diffusion tensor imaging (DTI) have a well-established role in the radiological assessment of brain tumours. The utility of these diffusion-based sequences in staging remains suboptimal, however [4]. Both DWI and DTI assume that the diffusion of water molecules involves random, Brownian motion. Following this assumption the probability distribution function (PDF), the chance of a proton diffusing between two points in a given time is thought to follow a Gaussian distribution [5]. The apparent diffusion coefficient (ADC) is based on the standard deviation of this PDF and DTI extends this by deriving the ADC in a direction-dependent manner [5, 6]. Although the Gaussian model in DWI and DTI holds true for pure liquids, it overlooks the in vivo effect of the complex cytoarchitecture of organic tissue formed of various compartments, cell types and intracellular constituents. The true PDF, therefore, exhibits non-Gaussian behaviour and the way this deviates from a Gaussian PDF can be assessed using the dimensionless statistical measure called kurtosis. Diffusion kurtosis imaging (DKI) is a novel extension of DTI and provides the degree of directional, non-Gaussian diffusion, i.e. the diffusion kurtosis tensor [7, 8]. Although the actual physiological basis of DKI remains unclear, the notion is that microstructural variances between gliomas of different grades will result in different DKI parameters, e.g. mean kurtosis (MK), and therefore will potentially provide a more accurate, non-invasive biomarker for glioma staging [5, 9]. Early research is encouraging. Two prior meta-analyses which looked at the diagnostic accuracy of DKI for glioma grading projected a pooled area under the curve (AUC) of 0.94 and 0.96 for MK [10, 11]. We attempted to consolidate the preliminary meta-analyses evidence in the topic through an updated systematic review and meta-analysis. These earlier studies also only analysed the ability of DKI to differentiate glioma grade with no probing of DKIs role in differentiating glioma from other intra-axial tumours. This is an important question as it assesses the real-world applicability of DKI as a non-invasive imaging tool in tumours of an unknown lineage that have not been sampled. To address these issues, our systematic review and meta-analysis will scrutinise two research questions; firstly, we will further assess the diagnostic accuracy of DKI in differentiating low-grade glioma (LGG) from high-grade glioma (HGG) by broadening the inclusion criteria and including recent studies that adhere to the up-to-date WHO 2016 glioma classification and include IDH genotyping. To answer this first question, we will specifically look at the mean difference in MK between HGG and LGG and the overall diagnostic accuracy of DKI. Secondly, for the first time, we will also review the role of DKI in differentiating gliomas from other intra-axial tumours.

A comprehensive review protocol was set up according to the Preferred Reporting Items for Systematic Review and Meta-analysis Protocols (PRISMA-P) statement and guidance from the Joanna Briggs Institute Reviewers’ Manual for the systematic review of studies of diagnostic test accuracy [12, 13]. The protocol was registered in the international prospective register of systematic reviews, PROSPERO (registration number: CRD42018099192) and details have been published previously [14]. For the development of this full systematic review, the Preferred Reporting Items for a Systematic Review and Meta-analysis of Diagnostic Test Accuracy Studies (PRISMA-DTA) checklist was used [15]. A detailed PRISMA checklist is available in Online Resource 1.

To date, there have been only two published meta-analyses which have looked at the diagnostic test accuracy (DTA) of DKI in glioma discrimination and both demonstrated very promising initial results [10, 11]. Our systematic review and meta-analysis extend this further by increasing the number of electronic databases, removing language restrictions and broadening the eligibility criteria. Importantly regarding DTA of DKI in addition to studies, which follow the WHO 2007 classification, we also included studies using the current WHO 2016 classification, the latter the IDH genotype. Our systematic review included a total of 19 studies, and for looking at the role of DKI in glioma stratification, two meta-analyses were performed. For mean difference analysis of mean kurtosis (MK) across 12 studies, we demonstrated a statistically significant mean difference in MK between high-grade gliomas and low-grade gliomas of 0.22 (95% CI 0.19–0.25). These results are comparable with the Delgado et al., where across 10 studies, they found a significant MK mean difference between high- and low-grade gliomas of 0.17 (95% CI 0.11–0.22) [10]. Our second meta-analysis included 9 studies, assessing the overall diagnostic test accuracy of DKI in grading gliomas, further confirmed its high diagnostic potential with an 87% sensitivity, 85% specificity and 0.92 pooled area under the curve. This is similar to the Delgado group findings which were across a smaller group of 5 studies, which produced a pooled sensitivity and specificity of 85% and 92% respectively and a pooled area under the curve of 0.94.

Looking at heterogeneity, the Delgado et al. subgroup analysis found the type of astrocytoma, the maximum b value and repetition time as significant study level characteristics moderating the diagnostic performance of DKI. Our larger study similarly showed moderate heterogeneity in study scanning parameters (such as TE, TR, max b value, number of diffusion directions). However, across a wider range of studies and by using a bivariate meta-regression model, we found no significant impact of the different study technical characteristics on the diagnostic accuracy. This finding is particularly encouraging as it suggests generalisable clinical utility. Nevertheless, optimization and standardisation of these parameters and post-processing techniques are still required to enable multi-centre quantitative studies, a key concept in generating higher-level evidence. It was not possible to determine a definite MK cutoff value for differentiating HGGs from LGGs. Despite this limitation based on the range of thresholds used in each separate study, we can extrapolate that the optimal cutoff value for MK is placed between 0.5 and 0.6.

Of the 19 studies in the systematic review, 4 studies looked into DKIs role in stratifying IDH mutation status as per the 2016 WHO classification [9, 20, 34, 35]. These studies reported significantly different MK values between IDH wild type and mutant suggesting its potential role as a surrogate marker for IDH phenotyping. Unfortunately, all 4 studies were from the same institute and based on an overlapping dataset preventing a subgroup meta-analysis [9, 20, 34, 35]. Recently, however, after the date this systematic review search was performed, two studies have been published comparing the diagnostic performance of DKI and DTI in predicting the IDH mutation status in gliomas [36, 37]. Both groups considered MK and mean diffusivity (MD) as the main DKI and DTI parameters respectively and concluded that MK can identify IDH mutation status with higher diagnostic value than MD [36, 37].

For the second question, looking at the role of DKI in differentiating glioma from non-gliomatous CNS tumours, our literature search only identified two studies which did not allow us to obtain any conclusive results [18, 19]. These studies showed however encouraging results that need to be reproduced in the future. Yan Tan et al. analysed MK values in solid tumour parts and the periphery of high-grade astrocytomas (HGAs) and solitary metastatic lesions, concluding that MK values differed significantly in the periphery between the two entities. MK values were also more sensitive than diffusion tensor imaging (DTI) metrics [19]. Using DKI, Pang et al. aimed to differentiate between HGGs and primary CNS lymphomas (PCNSLs) [18]. They reported significantly higher MK in PCNSLs than HGG, which could perhaps be explained by the hypercellular nature of lymphomas microenvironment.

Our study has a few possible limitations. Several of the studies are of limited sample size; none of the included studies reported individual patient data and it is not possible to account for differences in post-processing techniques like tumour ROI. Finally, the bivariate meta-regression analysis lacked a multivariate assessment.

Our work further confirms that DKI has a very good diagnostic performance in stratifying high- and low-grade gliomas. The consistent accuracy across different studies with varied acquisition and post-processing techniques importantly also implies that DKI is a technique that may be generalisable and clinically useful across different institutions and populations. Optimisation and standardisation of DKI techniques are still needed however to ensure consistency and parity. We also show its potential role as a surrogate marker for IDH phenotyping, although this requires further investigation. Finally, this study highlights the need to further explore the role of DKI in characterising non-glioma tumours.