Discrimination of epileptogenic lesions and perilesional white matter using diffusion tensor magnetic resonance imaging

Abstract

The aim of this study was to evaluate whether ganglioglioma (GGL), dysembryoplastic neuroepithelial tumour (DNET) and FCD (focal cortical dysplasia) are distinguishable through diffusion tensor imaging. Additionally, it was investigated whether the diffusion measures differed in the perilesional (pNAWM) and in the contralateral normal appearing white matter (cNAWM).

Six GGLs, eight DNETs and seven FCDs were included in this study. Quantitative diffusion measures, that is, axial, radial and mean diffusivity and fractional anisotropy, were determined in the lesion identified on isotropic T2 or FLAIR-weighted images and in pNAWM and cNAWM, respectively.

DNET differed from FCD in mean diffusivity, and GGL from FCD in radial diffusivity. Both types of glioneuronal tumours were different from pNAWM in fractional anisotropy and radial diffusivity. For identifying the tumour edges, threshold values for tumour-free tissue were investigated with receiver operating characteristic analyses: tumour could be separated from pNAWM at a threshold ≤ 0.32 (fractional anisotropy) or ≥ 0.56 (radial diffusivity) *10^–3 mm²/s (area under the curve 0.995 and 0.990 respectively).

While diffusion parameters of FCDs differed from cNAWM (radial diffusivity (*10^–3mm/s²): 0.74 ± 0.19 vs. 0.43 ± 0.05; corrected p-value < 0.001), the pNAWM could not be differentiated from the FCD.

Introduction

Hippocampal scleroses, long-term epilepsy-associated tumours (LEATs), ganglioglioma (GGL) and dysembryoplastic neuroepithelial tumours (DNETs), as well as focal cortical dysplasia (FCD), are the most common subtypes of epileptogenic lesions in epilepsy surgery candidates.¹ Epilepsy surgery aims at resecting the epileptogenic zone, which may include perilesional tissue. Based on the assumption that diffusion measures mirror the microstructure,² we sought to investigate whether diffusion tensor imaging (DTI) is able to identify the epileptogenic zone, specifically whether it is able to demonstrate alterations within perilesional tissue.

Methods

Patients

This retrospective study was approved by the institutional ethics committee (283/15).

All patients were recruited in southern German epilepsy centres and have a European ethnic background. They underwent epileptic surgery for suspected FCD or LEAT at the Neurosurgical Department of the University Medical Centre Freiburg between 2011 and 2016 and were screened for eligibility. Inclusion criteria were: a) availability of preoperative volumetric high-resolution, contrast-enhanced T1w and T2- or FLAIR-weighted imaging, b) availability of 61-direction diffusion imaging with two b-values (b = 0 s/mm²; b = 1000 s/mm²). As DTI was performed only during routine preoperative imaging in the case of eloquent lesions, for example, adjacent to the optic radiation or corticospinal tract, 21 patients were ultimately enrolled into analyses. Patients with mesiotemporal sclerosis, hippocampal atrophy, previous temporal lobe surgery or high-grade brain tumours were excluded to ensure a homogenous cohort.

The final cohort (N = 21, 13 females) included seven FCDs (five type IIA and two type IB), six GGLs WHO °I and eight DNETs WHO °I, confirmed by definitive histopathology. Median age was 18.2 years (range 7–68 years). An overview of the patients’ clinical and histopathological data and outcome is given in Table 1. All patients were studied on a 3 Tesla magnetic resonance imaging (MRI) scanner with an epilepsy-dedicated protocol including a 3D isotropic T2- or FLAIR- sequence (voxel size 1 mm³). In addition, a 61-direction diffusion imaging (DTI) with b-values of 0 s/mm² and 1000 s/mm² (voxel size 8 mm³) was used.

Table 1.

Overview of the patients’ clinical and histopathological data and outcome.

Patient	Sex	Pathology	Lesion localization	Age at epilepsy onset (years)	Age at operation (years)	Outcome Engel-classification
1	Male	FCD lla	Left occipitolateral	7	34	IIA
2	Male	FCD lla	Left frontal	2	51	IIIA
3	Female	FCD lla	Left frontal	1	5	IA
4	Male	FCD lla	Right parietal	13	66	IIA
5	female	FCD lla	Left frontal	1	15	IB
6	Female	FCD llb	Right parietal	11	14	N/A
7	Male	FCD llb	Left temporobasal	8	25	IB
8	Male	Ganglioglioma °I	Left temporomesial	8	15	IIA
9	Female	Ganglioglioma °I	Left temporomesial	10	11	IA
10	Female	Ganglioglioma °I	Left temporolateral	10	10	N/A
11	Female	Ganglioglioma °I	Left temporomesial	35	35	N/A
12	Female	Ganglioglioma °I	Left temporomesial	20	21	IA
13	Female	Ganglioglioma °I	Left occipitolateral	5	6	IIB
14	Female	DNET	Right frontal	13	41	IIIA
15	Female	DNET	Right occipital	5	13	IA
16	Male	DNET	Right frontal	12	16	IIB
17	Male	DNET	Left perietal	12	14	IA
18	Female	DNET	Left occipital	21	21	N/A
19	Male	DNET	Right temporomesial	41	56	IA
20	Female	DNET	Left temporobasal	13	14	IA
21	Female	DNET	Right temporolateral	30	42	N/A

FCD: focal cortical dysplasia; DNET: dysembryoplastic neuroepithelial tumour

Imaging methods

The following pre-operative datasets had been obtained: T2- (TE 231 ms, TR 2500 ms, 1 × 1 × 1 mm³) or FLAIR (TE 244 ms, TR 6000 ms, TI 2000 ms, 1 × 1 × 1 mm³) SPACE sequences and a 61-direction DTI with b-values of 0 s/mm² and 1000 s/mm² (TE 80 ms, TR 3100 ms, TI 90 ms, 2 × 2 × 2 mm³). All images were taken as part of the standard epilepsy protocol at our institution, except for DTI being performed only in the case of a lesion’s eloquent location (i.e. near the visual radiation or the pyramidal tract). Imaging was performed using a 3-Tesla Magnetom™ Tim® Trio (Siemens, Erlangen, Germany) and a 12-channel head coil.

Data processing

DTI datasets were co-registered to isotropic T2- or FLAIR datasets using SPM 8 (Statistical Parametric Mapping, Wellcome Department of Imaging Neuroscience, Institute of Neurology, University College London, UK) based on Matlab 8.0.0.783 (The MathWorks Inc., Natick, MA, USA). Regions of interest (ROIs) based data processing was performed with the in-house platform NORA (www.nora-imaging.com). ROIs were delineated by authors IM (26 years of experience) and NL (eight years of experience). One ROI comprised the lesion identified on 3D-T2 or FLAIR imaging. Special care was taken to avoid cysts, cerebrospinal fluid (CSF) and vessels, while trans-mantle dysplasia and blurring were included in the ROI. The larger voxel size of the DTI datasets in comparison with conventional MRI was taken into account in the ROI generation.

To assess degrees of perilesional abnormality, another ROI was drawn as an onion-like layer into the perilesional normal-appearing white matter (pNAWM). Corresponding areas in the contralateral white matter were identified, so that a mirror-image comparison structure (contralateral normal-appearing white matter (cNAWM)) was created. Mean, axial and radial diffusivity, as well as fractional anisotropy, were obtained from all ROIs.

Statistics

Statistics were performed with SPSS 23.0.0.0 (IBM Corp., Böblingen, Germany).

First, a Lilliefors test for normal distribution was performed. The normally distributed diffusion parameters were analysed with an analysis of variance followed by least significant difference and Tamhane post-hoc testing. Not normally distributed parameters were analysed using the Mann–Whitney U test. This applied to the fractional anisotropy of the ROIs’ T2-/FLAIR-hyperintensity and pNAMW.

p-values were corrected for multiple testing according to Bonferroni–Holm and a significance level of p <0.05 was set. Furthermore, a receiver operating characteristic (ROC) analysis was performed to identify fractional anisotropy and radial diffusivity thresholds with sufficient sensitivity and specificity to distinguish the lesion from its pNAWM and include the gross tumour mass. Standard error was determined under non-parametric assumptions.

Results

In Figure 1, a panel of structural images illustrating the lesions and the image quality is given.

Figure 1.

Panel of representative images. In (a) to (c) examples for focal cortical dysplasias type IIA ((a) and (b)) and IB (c) are given. In (d) to (f) gangliogliomas WHO °I, and in (g) to (I) dysembryoplastic neuroepithelial tumours are displayed. Arrows indicate all pathologies.

Differentiation of the lesions

GGLs and DNETs could not be distinguished using diffusion measures. DNETs, however, had a significantly higher mean diffusivity compared with FCDs (corrected p-value (p_corr) = 0.022). For detailed values see Table 2. GGLs showed a significantly higher radial diffusivity than FCDs (p_corr = 0.046).

Table 2.

Diffusivities and fractional anisotropy of the lesions and their respective areas of normal appearing white matter.

		AD		RD		MD		FA
	n	Mean	SD	Mean	SD	Mean	SD	Mean	SD
T2/FLAIR hyperintensity
FCD	7	1.00	0.12	0.74	0.19	0.91	0.15	0.21	0.09
GGL	6	1.22	0.28	1.01	0.31	1.14	0.29	0.13	0.05
DNET	8	1.41	0.45	1.19	0.46	1.34	0.45	0.13	0.08
pNAWM
FCD	7	0.94	0.08	0.54	0.14	0.80	0.09	0.37	0.09
GGL	6	0.93	0.09	0.49	0.13	0.77	0.09	0.38	0.09
DNET	8	0.89	0.08	0.45	0.05	0.72	0.04	0.40	0.04
cNAWM
FCD	7	0.89	0.07	0.43	0.05	0.74	0.04	0.42	0.08
GGL	6	0.86	0.05	0.42	0.07	0.71	0.05	0.42	0.07
DNET	8	0.88	0.11	0.39	0.05	0.69	0.04	0.49	0.07

Means ± standard deviations of axial (AD), radial (RD) and mean diffusivity (MD) given in [*10^–3 mm²/] and fractional anisotropy (FA) [dimensionless]. The significant differences discussed in the main text are shown by pairs of the same colour. The corresponding corrected p-values (p_corr) are: p_corr = 0.046, p_corr = 0.022, $\mathit{skyblue}1p_{\mathit{corr}}=0.04$, $\mathit{orange}{p}_{\text{corr}}<0.001$, $\mathit{yellow}{p}_{\text{corr}}<0.001$, $\mathit{skyblue}{p}_{\text{corr}}<0.001$, $\mathit{green}{p}_{\text{corr}}<0.001$, Lesion vs. lesion is marked with bold face coloured characters.

FCD: focal cortical dysplasia; GGL: ganglioglioma; DNET: dysembryoplastic neuroepithelial tumour; pNAWM: perilesional normal appearing white matter; cNAWM: contralateral normal appearing white matter.

Lesion versus pNAWM

In GGLs, the T2-/FLAIR-hyperintense area had a significantly higher radial diffusivity and a significantly lower fractional anisotropy than pNAWM (p_corr < 0.001, p_corr = 0.04, respectively). For DNETs, the T2-/FLAIR-hyperintensity differed from the pNAWM by higher radial diffusivity and decreased fractional anisotropy (p_corr < 0.001). In contrast, for FCD a distinction based on diffusivities between the T2-/FLAIR- hyperintensity and pNAWM was not possible. Interestingly, the radial diffusivity of the FCDs’ T2-/FLAIR-hyperintense area was significantly higher than the contralateral anatomically analogous white matter (cNAWM), p_corr < 0.001.

As the mean values of fractional anisotropy and radial diffusivity of the T2-/FLAIR-hyperintense area and the pNAWM did not differ statistically in both tumour entities, they were combined into one group, and a ROC analysis was performed to obtain a threshold, above which the fractional anisotropy or the radial diffusivity value was expected to belong to tumour or perilesional white matter; see also Figure 2. The ROC analysis of the fractional anisotropy revealed an area under the curve (AUC) of 0.995 (p_corr < 0.001; 95% confidence interval (CI): 0.978–1.0): With a fractional anisotropy threshold of 0.32 it was possible to identify all tumours (DNETs and GGLs) with a sensitivity of 100% and a specificity of 93% (95% CI: 0.785–1.0). The mean fractional anisotropy of the tumours is 0.13 (± 0.06) compared with the mean fractional anisotropy of the pNAWM of 0.39 (± 0.06, p_corr < 0.001). The ROC analysis of the radial diffusivity revealed an AUC of 0.990 (p_corr < 0.001; 95% CI: 0.963–1.0): With a radial diffusivity of 0.56, all tumours (DNETs and GGLs) could be reliably identified with a sensitivity of 100% and a specificity of 93% (95% CI: 0.785–1.0) (mean ± SD: tumour 1.11 ± 0.39 * 10^–3 mm²/s vs. pNAWM 0.54 ± 0.14 * 10^–3 mm²/s, p_corr < 0.001).

Figure 2.

Receiver operating characteristic (ROC) analysis of radial diffusivity (red) and fractional anisotropy (blue) for the differentiation between tumour and perilesional normal appearing white matter. The ROC analysis of the fractional anisotropy revealed an area under the curve (AUC) of 0.995 (corrected p-value (p_corr) < 0.001): With a fractional anisotropy of 0.32 it is possible to identify all tumours (dysembryoplastic neuroepithelial tumours (DNETs) and gangliogliomas (GGLs)) with a sensitivity of 100% and a specificity of 93% (confidence interval (CI): 0.785–1.0). The ROC analysis of the radial diffusivity revealed an AUC of 0.990 (p_corr < 0.001): With a radial diffusivity of 0.56, all tumours (DNETs and GGLs) can be reliably identified with a sensitivity of 100% and a specificity of 93% (CI: 0.785–1.0).

The assumption in choosing those thresholds was that using these the whole tumour mass was included in the generated ROI, thus granting a high sensitivity.

The radial diffusivity value of pNAWM (0.54 ± 0.14 * 10^–3 mm²/s) did not differ with significance from FCDs (0.74 ± 0.19 * 10^–3 mm²/s) and their cNAWM (0.43 ± 0.05 * 10^–3 mm²/s), although being in an intermediate position.

pNAWM versus cNAWM

To answer the question whether perilesional white matter is pathologically altered, ROIs of pNAWM and cNAWM were compared. No difference between ipsilateral and contralateral pNAWM was found, neither in the LEATs nor in the FCDs.

Discussion

Lesion versus pNAWM

Especially noteworthy are the identified thresholds differentiating GGL and DNET from their pNAWM. These thresholds can be used to define borders of resection to include the whole tumour mass with a high sensitivity and specifity. In both the fractional anisotropy and the radial diffusivity, the threshold with the sensitivity of 100%, a relatively lower specifity has been chosen because the resection of false negative tissue would generally be more tolerable than not removing the gross tumour mass. Of course, this would have to be adapted to the lesion’s eloquent location.

This means that with an AUC of 0.995, a fractional anisotropy of ≤ 0.32 certainly would include the tumour. Radial diffusivity with an AUC of 0.990 is also capable of differentiation. For the radial diffusivity, a value ≥ 0.56 * 10^–3 mm²/s distinguishes the lesion from its pNAWM.

This is of great importance, as the outcome is largely dependent on the gross resection of the tumour. The usage of the given thresholds can improve the approach in surgical therapy by preoperatively identifying the gross tumour mass with an objective value. It also could be applied to machine learning approaches on tumour identification.

For FCDs, it is also assumed that the epileptogenic zone includes the MRI-visible lesion and perilesional tissue. This study, however, could not distinguish any threshold for the edge of the FCD as the diffusion measures in the MRI-visible lesion and the perilesional tissue were not different. This may indicate the blurred borders of FCD and the adjacent healthy matter.

Differentiation of lesions from each other

Differentiation of GGL and DNET is of importance because of the different prognosis of the entities and different approaches to the gross total resection and is therefore extremely relevant for clinical practice as the general surgical strategy needs to be adapted to the entity.

We did not detect any difference between the diffusion measures of GGL and DNET in any ROI, consistent with previous studies in the literature. The reason could be the microscopic substructure’s similarity with similar proportions of intracellular and extracellular space in both entities, as diffusion imaging represents the microstructural level of cell size and orientation of cellular structures.³

Whereas GGLs and DNETs were generally indistinguishable in terms of diffusion parameters, they could be differentiated from FCDs through one diffusion parameter: a higher mean diffusivity of DNETs allowed differentiation from FCDs; a higher radial diffusivity separated GGLs from FCDs.

The possible differentiation between FCDs and the LEATs in diffusion imaging is consistent with their histopathological representation. FCDs with predominantly neuronal character clearly differ from glioneuronal tumours,^4,5 showing an increased demyelination and axonal degeneration. This may lead to typical diffusion measure alterations; that is, decreased fractional anisotropy and increased mean and radial diffusivity.⁶

pNAWM versus cNAWM

Numerous studies indicate a role of the perilesional tissue in epileptogenesis and emphasize the importance of pre-surgical determination of the epileptogenic zone.⁷ Possible causes are alterations of cytoarchitecture and neurochemistry in the pNAWM.^8,9 The preoperative identification of the gross epileptogenic zone is of importance as it can affect the extent of resection.

In LEATs, no significant difference between pNAWM and cNAWM was detected in the diffusion measurements. In our opinion, this indicates that the perilesional white matter is normal at a distance of about 2 mm from the edge of the lesion (safety margin due to voxel size). This is consistent with the literature, where a perilesional zone of 1–2 mm was reported.¹⁰ This information can be taken into account in the planning of surgery extent.

For FCD, the situation is different. In our study, no diffusion parameter was found to distinguish the MRI-visible lesion from its pNAWM, nor the pNAWM from its cNAWM. In the literature three FCDs with T2-hyperintensity were reported to have a change of more than two standard deviations in apparent diffusion coefficient (ADC) of the perilesional compared with the healthy tissue, whereas two FCDs without T2-hyperintensities revealed no changes.¹¹ Possibly this difference between the reported differences lies in the type of FCD.¹¹ The lack of distinction in our study between FCDs and pNAWM as well as pNAWM and cNAWM, while a difference between FCDs and cNAWM was detected, may indicate gradual differences between these regions. In fact, the radial diffusivity value of pNAWM was actually in an intermediate position between FCDs and their cNAWM. So it should be an object of further investigation whether a differentiation of the FCD and the pNAWM is possible to identify the epileptogenic zone in FCD.

Limitations

A strong limitation of this study is the small sample size. But the significant results (by using conservative corrections for multiple testing) speak for striking differences between the diffusion measures. The literature about these entities is limited and there is no publication where all entities are compared within one study. Moreover, only values for fractional anisotropy and ADC are reported in the literature.

When comparing our results with the literature, one has to take into account that the ADC – being averaged by three diffusion directions – is only approximately similar to the mean diffusivity (calculated as the mean of the three eigenvalues) due to different effects of the measurement noise and the higher-order terms in the diffusion-weighted signal (kurtosis, etc.). Therefore, both measures are considered here for comparison with the literature, although it is only an approximation.

For GGL, ADC values of > 1.0 *10^–3 mm²/s, 1.36 *10^–3mm²/s (mean value) and 1.45 *10^–3 mm²/s (minimal value) were reported.^12–14 Our own mean diffusivity value is within the lower range of these reports. For DNET, only two ADC values were described in the literature, ranging from 2.55 = to 2.6 *10^–3 mm²/s.^15,16 This value is different from our mean diffusivity value of 1.34 *10^–3 mm²/s. For the FCD, two ADC values were reported, ranging from 0.89 to 2.8 *10^–3 mm²/s.^17,18 Our own mean diffusivity value lies in the lower range of the reports. The ADC values of NAWM were 0.8–0.9 * 10^–3 mm²/s in the literature,^6,7,9 whereas mean diffusivity of cNAWM in our study ranged from 0.69 to 0.74 *10^–3 mm²/s for the three entities. The lower values in our study may have several causes. First, there is no study comparing all three entities with each other in one study. This leads to a large variance in the methods of the groups in the literature. Second, in our study the inclusion of CSF and vessels was strictly avoided, as was the inclusion of tumour cysts. Even if in some studies in the literature this was mentioned,⁷ there might have been a different way to account for CSF or cysts, leading to consistently lower values of all diffusion measures in our study. Not taking the different voxel size into account could lead to false high results, so we consequently avoided fluid compartments.

Conclusion

Fractional anisotropy and radial diffusivity allow the differentiation of pNAWM from glioneuronal tumours. FCDs can be distinguished from DNET by a lower mean diffusivity, and from GGL by a lower radial diffusivity.

Funding

This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

Conflict of interest

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.