Types of studies

We included cross‐sectional studies with retrospective or prospective design, and excluded case‐control studies. Non‐English literature studies were eligible for inclusion.

Participants

We considered studies that enrolled paediatric and adult patients with solid and non‐enhancing infiltrative tumours on standard contrast enhanced MRI and who had subsequent WHO grade histologic classification following biopsy or resection. The interval period between MR perfusion and histologic diagnosis must be within two months, as LGGs can transform into HGGs over that period of time. As this review focused on primary and not recurrent disease, participants who had received previous surgery (either biopsy or resection) or oncological treatment (chemotherapy or radiotherapy) prior to the index test were excluded.

Index tests

Studies that employed dynamic susceptibility (DSC) and/or dynamic contrast‐enhanced (DCE) MR perfusion were considered for the review. They were included when data on rCBV and/or K^trans were provided, regardless of the method of acquisition, imaging analysis or post‐processing performed. Studies that used arterial spin labelling MR perfusion were excluded, as this method has been largely utilised as a research tool.

For DSC MR perfusion, we adopted a common rCBV threshold of 1.75 following the landmark study of Law and colleagues (Law 2003). This value is widely accepted and is used and recommended in clinical practice (Al‐Okaili 2006). In this review, a positive test or rCBV < 1.75 implied an LGG diagnosis while a negative test or rCBV > 1.75 suggested an HGG diagnosis (Appendix 1). With reference to the seven studies included in this review, five studies did not report an rCBV threshold, while the other two studies used or reported various rCBV thresholds (1.29, 1.5, 5.66).

For DCE MR perfusion, there is no consensus in the field regarding a threshold. Only one out of seven included studies reported K^trans and did not use a threshold.

The generation of MR perfusion data is a multistep process involving acquisition, post‐processing and imaging analysis. Technical differences may arise at each step, which can ultimately influence the perfusion value obtained. Thus it is acknowledged that there may not be an appropriate single threshold when using multicentre MR perfusion data.

Target conditions

The target condition is LGGs (disease positive in quantitative analysis), with the alternative condition being HGGs (i.e. non‐LGG or disease negative in quantitative analysis). Studies must have both conditions to be included in the review.

We included studies that enrolled WHO Grade II diffuse astrocytomas, oligodendrogliomas and oligoastrocytomas, their Grade III anaplastic counterparts and glioblastomas (see Table 1). Other WHO Grade II‐IV gliomas of different histologic type were excluded, because they are rare and usually have distinct clinical characteristics. Grade I tumours were also excluded because their distinct clinical and radiological features separate their management from that of grade II tumours. In this review, gliomas were included irrespective of patient age and tumour location, although it is recognised that paediatric and adult gliomas, and supratentorial and brainstem gliomas may be histologically similar but have biologic and molecular differences (Paugh 2010).

Tumour histologies outside of these target conditions but having similar solid, non‐enhancing appearance as LGGs were recorded and excluded from the analysis.

Reference standards

The reference standard is histological diagnosis assessed according to the WHO 2007 criteria (Table 1). We recorded the method of tumour sampling (biopsy or resection) per study. Reference test failures were treated as no histologic confirmation, and excluded from the analysis.

There are two potential limitations with histology. First, histological diagnosis can be subjective and suffer from significant intra‐ and inter‐observer variability, even among neuropathologists (Brat 2008). Recently, the WHO classification (Louis 2016) has been substantially updated to incorporate molecular profiles in the diagnosis of diffuse gliomas, providing a layer of objectivity to tumour evaluation. However, these are not always available or reported, and were not used in this review. Second, tissue sampling is constrained by the surgical technique of tissue acquisition. Because gliomas are morphologically heterogeneous, they are susceptible to sampling errors, particularly in biopsies. Small volumes of tissue may not be representative of the most malignant component and are therefore less ideal compared with full tissue evaluation afforded by maximal resections. In fact, agreement in histological typing and grading between biopsies and resections has been found to be low (Muragaki 2008).

Electronic searches

We performed a systematic search for literature with quantitative data. Articles were retrieved from the following electronic databases.

1. MEDLINE via OvidSP (Appendix 2, 1996 to 9 November 2016).

2. Embase via OvidSP (Appendix 3, 1996 to 9 November 2016).

3. Web of Science Core Collection, specifically Science Citation Index Expanded and Conference Proceedings Citation Index ‐ Science via Thomson Reuters Web of Science (Appendix 4 1990 to 9 November 2016).

The search was restricted to studies in humans, and was limited to studies in the last 20 years, since MR perfusion is a relatively recent MRI technique.

Searching other resources

Additional references were identified by manually searching the references of relevant review articles. We also used the 'related articles' feature in PubMed to identify articles similar in keywords and database subject headings to the original included studies.

We sought unpublished studies by searching the conference proceedings, available online, of the following radiologic and neuroradiologic societies (years accessed).

1. The Radiology Society of North America (2003 to 2016).

2. The American Society of Neuroradiology (1996 to 2016).

3. The European Society of Radiology (1999 to 2016).

4. The European Society of Neuroradiology (1995 to 2016).

Selection of studies

Prior to abstract review, two review authors (JMA, MGH) performed preliminary screening of titles to exclude irrelevant references. Subsequent screening was carried out by at least two authors (JMA, DMF, MGH, EKCL or JMP). Disagreements were resolved by consensus or third person.

During the abstract review, potentially eligible publications were selected based on the following criteria.

1. Did the study include participants suspected to have primary infiltrative gliomas (including LGG)?

2. Were brain tumours confirmed by histology?

3. Was MR perfusion performed?

An initial review of full‐text publications was then conducted to establish that they meet the following criteria.

1. Did the study enrol participants with solid non‐enhancing brain tumours on standard contrast‐enhanced MRI scan?

2. Were measurements on rCBV and/or K^trans given?

3. Can a two‐by‐two table be extracted from the provided data?

When this information was not readily available from the published data, we contacted primary authors for clarification. In particular, information on the MRI appearance of the gliomas, i.e. if they were solid and non‐enhancing, and were not always explicitly reported in studies. We also requested for patient‐level data from authors to allow selection of tumours relevant to our review. The data included the individual perfusion values, the method of histologic confirmation, and interval period between the index test and histologic confirmation.

Detailed data extraction of full‐text paper or conference abstract was carried out to determine final eligibility based on the inclusion criteria (see Criteria for considering studies for this review).

Data extraction and management

Two review authors (JMA, DMF) independently completed a data extraction form on all included studies. The following data were retrieved.

1. General information: title, journal, year, publication status, and study design (prospective versus retrospective).

2. Sample size: number of participants included, those meeting the target and alternative conditions (i.e. other‐gliomas and non‐gliomas), and reference test failures.

3. Study setting: country and type of hospital.

4. Baseline characteristics: age, sex.

5. Clinical reference standard test: method of tissue sampling (biopsy or resection), blinding from MR perfusion results, interval period between MR perfusion and tissue diagnosis.

6. Histologic diagnosis and WHO grade of tumour. In this review, LGGs are considered disease positive, and HGGs are disease negative.

7. Details of the index test, including the strength of the MRI scanner, method of MR perfusion performed (DCE or DSC), pulse sequence, use of contrast preload, post‐processing algorithm, imaging analysis, blinding from histological analysis if retrospectively processed.

8. Quantitative results of the index tests, including but not limited to rCBV, K^trans.

9. Authors' recommended threshold.

10. Number of true positive, false positive, true negative, false negative, and area under the receiver operating characteristic (ROC) curve, extracted based on authors' pre‐specified and recommended thresholds, and based on Law's threshold of 1.75. During the course of this review, we found that authors' pre‐specified and recommended thresholds were either missing or when available, heterogeneous, therefore we only used Law's threshold to construct the two‐by‐two contingency tables. A positive index test implied a diagnosis of LGG while a negative test suggested HGG. Thus, true positives are correctly diagnosed LGGs, false negatives are LGGs incorrectly labelled as HGG, true negatives are correctly diagnosed HGGs, while false positives are HGGs incorrectly labelled as LGG (Appendix 1).

We contacted primary authors via e‐mail for missing data and to obtain clarification on study methods.

Assessment of methodological quality

Each study was assessed for methodological quality using the revised Quality Assessment of Diagnostic Accuracy Studies tool (QUADAS 2) (Whiting 2011). QUADAS 2 assesses study quality through the use of signalling questions which facilitate assessment of risk of bias in four domains: patient sampling; index test; reference standard; and flow and timing. In addition, concerns regarding applicability were also rated in the first three domains. The QUADAS 2 tool for this review is described in Appendix 5 and was applied on study design and on the included patient data. To reduce uncertain risks, we sought clarification from authors when information was missing. e.g. blinding between MR perfusion and histology, details on the MRI acquisition and post‐processing.

Summative data from QUADAS 2 were used to describe the numbers of studies with high/low/unclear risk of bias as well as concerns regarding applicability. All studies with sufficient data were included in the general analysis, but studies with low risk of bias in the domains of reference standard and flow and timing were used in the sensitivity analysis.

Statistical analysis and data synthesis

The reported number of true positive, false positive, false negative and true negative cases were used to construct two‐by‐two tables based on the author's pre‐specified and recommended thresholds, and Law's threshold. When these counts could not be determined from published data, we contacted authors for individual patient‐level data.

Of the seven included studies, five studies did not report an rCBV threshold, while two studies pre‐specified or determined different rCBV thresholds (1.29, 1.5 and 5.66). One study used a pre‐specified threshold of 1.5, which was chosen as a mean between literature values (Law 2003; Zonari 2007). Thus, we used Law's threshold of 1.75 as a common threshold in the quantitative analysis for DSC perfusion, as this is more widely accepted in the field. Individual patient‐level data were classified using this common threshold to construct the two‐by‐two table per study.

Meanwhile, of the seven included studies, only one study reported K^trans results, precluding quantitative analysis for DCE MR perfusion.

Coupled forest plots showing pairs of sensitivity and specificity, with their 95% confidence intervals (CI), were constructed for each study using RevMan 5.3 (Macaskill 2010; Review Manager 2014).

To estimate the summary sensitivity and specificity, we applied the bivariate model (Reitsma 2005), which accounts for between‐study variability in estimates of sensitivity and specificity through the inclusion of random effects for the logit sensitivity and logit specificity parameters of the bivariate model. To generate the bivariate model parameters required to construct the SROC plot in Revman 5.3 (Review Manager 2014), the model was fitted using the lme4 package of R (R Core team 2013) following Partlett 2016.

Investigations of heterogeneity

As part of the secondary objective, we planned to investigate heterogeneity potentially arising from differences in participant characteristics and study design:

1. tumour histology (subgroups for astrocytomas, oligodendrogliomas, and oligoastrocytomas);

2. MRI strength (1.5T versus 3T);

3. MR perfusion technique (DSC versus DCE);

4. use of contrast preload;

5. post‐processing technique (arterial input function versus gamma variate fitting; region of interest versus histogram analysis);

6. method of histological sampling (biopsy versus resection).

Due to the insufficient number of studies, we were unable to perform regression approach to explore heterogeneity.

Subgroup analysis was conducted for tumour histology (astrocytomas versus oligodendrogliomas plus oligoastrocytomas) based on the bivariate model. As the data are sparse, the bootstrap method was used to compute the 95% confidence intervals for the estimates through the boot package of R (R Core team 2013). Other subgroup analyses were not possible due to the small number of studies and participants.

Sensitivity analyses

To determine the robustness of the outcome of the meta‐analysis, we conducted sensitivity analysis on good quality studies, defined as those with low risk of bias in domains 3 (reference standard) and 4 (flow and timing) as specified in the QUADAS 2 criteria (Appendix 5).

We also initially intended to conduct sensitivity analysis based on tumour status outside the target and alternative conditions (i.e. including other‐gliomas and non‐gliomas), but no data were available to perform this.

Assessment of reporting bias

We did not assess reporting bias as there are no clear‐cut methods to perform this for studies of diagnostic accuracy (Brazzelli 2009). It is noted that the index test under review is a widely available MRI technique and could be routinely employed in clinical practice (NCCN Guidelines 2016). As such, it would be unlikely to have a record of all attempted evaluations using this technique and therefore positive results are more likely to be reported in the literature (Irwig 1995). To minimise this, we included radiology and neuroradiology conference proceedings in our literature search, where negative results could be documented which may not have been published. Further, the small number of included studies was also inadequate to assess reporting bias by funnel plot (Deeks 2005).