Structural and functional MRI.

All patients underwent preoperative structural and functional brain MRI using a 3T scanner (Excite HDx, GE Medical Systems, Milwaukee, WI; eight channels head coils).

Standard diagnostic sequences included: 3D T2-FLAIR, T2*-GRE, T2-FSE and T1-SE. 3D T1-weighted fast-spoiled gradient echo (FSPGR) images of the whole brain (TR = 10.7 ms TE = 4.9 ms, matrix 256x256, flip angle 13°, slice thickness 1mm with no gap, FOV 25.6 cm, voxel resolution [1.0x1.0x1.0] mm³, acquisition time 5.60 min). Axial Diffusion Tensor Imaging (DTI) acquisition was performed with a single-shot EPI (Echo Planar Image) applying 30 different gradient directions uniformly distributed on a sphere through electrostatic repulsion with b-value = 0–1.000 s/mm² (TR = 16.000 ms, TE = 86 ms; BW = 250 kHz; voxel resolution: [2 x 2 x 2,4] mm³; FOV: 25.6 cm; 57 contiguous slices with 2.4 mm of thickness and no gap; acquisition time: 9.36 min).

All patients underwent a BOLD-functional MRI (EPI- GRE, TR/TE 2500/40ms, flip angle 90°, matrix size 128x128, FOV 24.0 cm, in-plane spatial resolution 1.8 mm, slice thickness 4 mm) with a block designed paradigm and 72 phase volumes obtained from 30 contiguous axial slices of 4mm/0gap thickness (acquisition time 193sec). Each period of the blocks (20sec rest - 20sec task) was repeated 4 times, with a final rest phase. The first epoch always lasted 4 phases (13s) more than the followings, not accounted for analysis, necessary to reach a steady-state condition.

The fMRI protocol included a string of three language tasks: Word Generation (WG) [28] and Rhyme Generation (RG) [29] tasks, for the assessment of phonemic fluency, and a comprehension task for the assessment of the semantic fluency [5]. Before starting the fMRI protocol, patients were instructed on the correct execution of each task outside the MR-room. During the WG task, patients wearing MR-compatible headphones covering both ears, listened to a neutral voice that pronounced a single randomly chosen letter every 5 seconds. During this time period, patients had to silently generate as many words they could that began with each listened letter. The RG task was built in the same way as the WG task, but patients were asked to listen to a short word (e.g. cane) and then silently produce a corresponding rhyming word (e.g. pane). For the semantic fluency task, patients were requested to listen to the description of a category of objects and silently think of as many as possible corresponding objects.

We excluded from the analysis fMRI acquisitions affected by head motion distortions and scanner artefacts. For fMRI BOLD data analysis, we used AFNI software [30]. We corrected images for slice acquisition timing differences and registered them to a reference volume in order to minimize movement-related effects. To perform a spatial smoothing operation, we used a three-dimensional Gaussian kernel with 3mm full width half maximum, in order to enhance signal to noise ratio and further reduce residual movement effects. We computed activation maps using a multiple regression analysis incorporating regressors and describing head movements, as obtained by the volume registration step. To model the hemodynamic response, we used a gamma variate function convolved with a boxcar wave function describing the temporal paradigm and evaluated the activation maps on the basis of a partial F-statistics and a cut-off at p<0.05 uncorrected [31]. As these maps could be affected by patients’ head movements and by global SNR variations, to be sure to select correct functional activated voxels, we visually analyzed the time course of each activated voxel to confirm the fitting with the applied paradigm.

Owing to the clinical characteristics of our sample (i.e. mainly children and overall a poorly collaborative cohort) we applied a subject-specific thresholding to the analysis of fMRI data (p values varied between 0.001 and 0.0001) in order to obtain BOLD maps as much as possible corresponding to the actual functional performance.

Twenty patients were included in the postoperative fMRI study, using the same protocol as for the preoperative mapping (the full MRI dataset is available as supporting information).

FMRI laterality index (LI).

In the subject-specific thresholded fMRI maps, we first selected, for each task of the individual patients, only voxels that showed a visual correlation with the applied time course in order to avoid the inclusion of voxels of pseudo-activations caused by noise or movement artefacts that the software hadn’t eliminated. Then, we grouped the selected voxels in four macro—regions of interest (ROIs), on the anatomical basis of simple theoretical functional language patterns: left and right anterior frontal ROIs to delineate the Broca’s area and the frontal dorsolateral region, and left and right posterior temporo-parietal-occipital ROIs to include the Wernicke’s area.

To compute the number of activated voxels in each ROI we used an automated post-processing script with AFNI software [32], according to the methodology described by Ruff and colleagues [33]. We calculated two types of Laterality Indexes, as follows: 1) an hemispheric index, globally comparing all the activated voxels of the left side (L) with those of the right side (R); 2) a functional-ROI index, comparing each functional ROI of L with those of R. To compute LIs, we applied the formula [L—R]/[L + R]. The values ranged from -1 to +1, with -1 indicating complete right hemisphere language dominance and +1 indicating complete left hemisphere dominance. Intermediate values reflected varying degrees of laterality. We set the LI value for defining atypical language lateralization at 0.20 (LI < 0.20 as atypical), according to literature [7, 19].

DTI.

Twenty-four (82%) out of 29 patients underwent also DTI analysis on arcuate fasciculus.

Diffusion MRI was processed using FSL tools. Head motion and image distortions induced by eddy currents in diffusion data were corrected using FSL’s eddy tool [34]. After distortion correction, DTI data were averaged and concatenated. Bayesian Estimation of Diffusion Parameters Obtained using Sampling Techniques was performed on diffusion data using the bedpostX tool of FSL [35, 36], which allows modelling of crossing fibers within each voxel of the brain. Subsequently, the probtrackX tool of FSL was used to perform tractography of left and right arcuate fasciculi for each subject. Regions of interest (ROI) were manually placed in the language areas that are directly connected by the long segment of the arcuate fasciculus [37]. Due to the variability in the location and size of pre-operative lesion/post-operative scar in the brains of our patients, in order to limit the reconstruction of spurious connections we chose to perform tractography using three anatomical ROIs instead of one or two as in the work by Catani and coworkers [37]. Fractional Anisotropy (FA) maps in the subject’s diffusion space, color-coded according to the principal diffusion direction V1, were used to draw ROIs in the posterior part of superior temporal gyrus (Wernicke’s area) and in the frontal lobe white matter (pars opercularis and triangularis, corresponding to Broca’s area). These two ROIs were defined as “seed” and “target” for tractography. Moreover, a “waypoint” mask was placed in the white matter underlying the inferior parietal cortex (Geschwind’s area). ROI tracing was performed by two neuroimaging experts blinded to clinical diagnosis. Reproducibility of ROI placement was ensured by careful visual inspection of the images.

To account for volume differences, the resulting tracts were normalized by the corresponding ‘waytotal’ value (i.e., the total number of streamlines surviving the tractography constraints, such as seed, target and waypoint masks), and a probability threshold of 0.01 was set to decrease the risk of erroneous findings due to spurious reconstructions. Finally, tracts were binarized.

After reconstruction of the arcuate fasciculus, we computed the volume of each tract mask and the histogram of the FA values in this mask for each side, normalized by the total number of voxels of the tract. Finally, we calculated the following histogram-derived measures:

• mean, median and mode (measures of central tendency);
• standard deviation (measure of variability);
• skewness (measure of asymmetry);
• kurtosis (can indicate the presence of outliers).

For each of the above-mentioned metrics (i.e., volume of the tract and histogram-derived measures), we calculated a laterality index according to the following formula: We set a cut-off value of 0.2 for determining the presence of asymmetry, based on previous reports [38, 39].

For the 20 patients who underwent postoperative functional MRI during language tasks, we calculated volume and FA changes of the arcuate fasciculus after surgery according to the following formula: with M being either volume or FA. Positive results indicated an increase of FA after surgery, while negative values indicated a decrease of FA after surgery. Pre- and post-operative values of the metrics were compared using paired t-tests with Bonferroni correction and statistical threshold at 0.05.

Finally, we defined whether language lateralization assessed through fMRI was related to possible asymmetries of the arcuate fasciculus in individual patients.

Neuropsychology.

We tested general cognitive abilities using standardized tools: GMDS-ER [40], Italian versions of the Wechsler Scales [41, 42] and measured non-verbal function using the Leiter-r Scale [43]. Since a limited population and its peculiar clinical features do not permit maximum homogeneity in a cognitive scale, we used a single psychometric measure (IQ/General Developmental Quotient). For the statistical analysis, we labeled patients as 0.1 or 2 respectively if they had normal-to-borderline cognitive development (IQ<70), mild delay (IQ 50–70) or moderate and severe delay (IQ<50).

We evaluated language function using tasks of naming and fluency and assessed the ability to name drawings of objects using the Boston Naming Test [44]. To assess verbal and semantic fluency, we used the word fluency test [45]. We classified language function performances in two categories: normal or pathological based on naming and fluency test scores.

To assess visual attention, each child was asked to perform a task of the barrage of target numbers among the target distractors. To measure the processing speed, we used a task from Wechsler scales (Digit Symbol) [41, 42]. As part of the assessment of verbal and working memory, we used two tasks as follows: digit span forward (SPANFW) and digit span backwards (SPANBW) derived from the Wechsler scales [41, 42].

Improvement, stability or deterioration of cognitive and language functions were defined based on possible changes of category according to the postoperative neuropsychology test scores.

Preoperative study.

For the purposes of the preoperative study, we hypothesized a causal model to test whether several clinical variables (possible predictors) would influence fMRI results (outcomes) and the latter (possible predictors) would in turn impact on cognitive performances (outcomes).

A causal model is a conceptual model that describes the causal mechanisms of a system, in this case the causal mechanisms of language reorganization in children and young adults with focal lesional epilepsy. The causal model provides clear rules for defining what independent variables need to be included or controlled for and what models to adopt [46]. Since the causal model defines the structure of the statistical analysis, we have not evaluated indirect associations and hierarchical structures of the data and models other than those defined by the casual model.

For each patient, we then analyzed the following three groups of variables, as follows:

1. Clinical features:
   1. General characteristics: gender, handedness, neurological deficits, age at seizure onset, seizure frequency, age at surgery, duration of epilepsy.
   2. Neuroimaging: localization, lateralization and extent (uni- or multilobar) of the epileptogenic lesion.
   3. Neurophysiology: topography of ictal and interictal EEG abnormalities (focal vs diffuse epileptiform activity i.e. spikes, polyspikes, spike and waves, sharp waves; ipsilateral vs contralateral epileptiform activity with respect to the epileptogenic lesion), extent of the SOZ (unilobar vs multilobar) and its relationship with the structural abnormality (whether superimposed to the lesion or extending beyond it).
   4. Surgical variables: Engel’s class, histopathology.
2. FMRI results:
   1. Intralesional vs perilesional activation and ipsilateral vs contralateral activation (with respect to the epileptogenic lesion) during each task. We classified activated voxels as “intralesional”, when included in the MRI visible structural abnormality, and “perilesional” when located in the apparently structurally normal brain bordering the lesion. When the limits of the structural abnormality were ill-defined at MRI (mainly in FCD type I), we anatomically referred to the surgical breach, providing that histopathology confirmed residual abnormal tissue bordering surgical excision margins.
   2. Global LI for each language task.
3. Neuropsychological findings:
   1. IQ (Intelligence Quotient) or GDQ (General Developmental Quotient) scores.
   2. Phonemic fluency and naming (pathological vs normal).
   3. Attention (pathological vs normal).
   4. Forward and backward span (pathological vs normal).

We defined a clinical cut off for the discretization of continuous variables wherever possible, because the discretization improves the efficiency of estimates when only a limited number of observations is available [47].

Owing to the high number of parameters of interest, we first performed an exploratory data analysis for each group through a factorial analysis for mixed data (FAMD) [48], in order to analyze the relationship between the individual parameters and select variables to be used as possible predictors in the subsequent analysis of associations [48]. FAMD is a principal component method allowing to explore the association between quantitative and qualitative variables and analyze the similarity between individuals in order to reduce a high number of variables in a few latent ones.

Then, for the analysis of associations, we used Pearson’s χ² test of independence and, if necessary, the exact test of Fischer, Student’s t-test and ANOVA (analysis of variance). To estimate the effect of each parameter on outcomes, we used generalized linear regression models.

We selected the covariates for the univariate analysis through the exploratory analysis. Owing to the small sample size, we did not consider possible confounders such as handedness in the univariate analysis, but we only estimated crude effects. If these effects were significant (p-value of regression coefficient < 0.05), we then computed multivariate models. For univariate and multivariate models, we used likelihood ratio test for overall model evaluation ad adjusted R² or pseudo R² as goodness of fit measures. For multivariate models, we carried out a sensitivity analysis for each model using a stepwise method and specified the significance level as addition to the model at 0.05; terms with p < 0.05 were eligible for addition. When an estimation was not possible in the univariate model i.e. GLR estimates had standard errors questionable or the confidence intervals cannot be estimated, we reported the exact test results. Given the exploratory nature of the study we did not perform any correction for multiple comparisons.

We carried out all statistical analyses using STATA version 13.0 (StataCorp. 2013. Stata Statistical Software: Release 13. College Station, TX: StataCorp LP) and R version 3.3.3 (2017-03-06) Copyright © 2017 The R Foundation for Statistical Computing.

Postoperative study.

In the 20 patients who underwent postoperative fMRI during language tasks, we computed LI changes after surgery (labelled as Delta-LIs). We performed an analysis of associations between Delta-LIs and the following clinical variables, which were chosen based on previous studies on the topic [49–54]: lesion lateralization (right vs left); postoperative scalp interictal EEG (normal vs pathological); histology (FCD, tumors, other); seizure outcome (Engel class IA vs Engel class IB-IV); disease and follow-up duration; IQ/GDQ scores, fluency and naming (improved, worsened, stable).

Functional MRI (Figs 1 and S1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Presurgical and postoperative 3T structural and functional MRI during WG task.

Right-handed 15-year-old girl with left frontal FCD IIb and left dominance for language. BOLD functional correlated signals (red-yellow colormap) are superimposed on axial FLAIR images (A, presurgical; B, postsurgical) and sagittal FLAIR sections (D, presurgical; E, postsurgical) across the lesion and the surgical breach (see white vertical lines). Structural MRI shows a left frontal malformed area located on a irregularly flat gyrus whose subcortical white matter is mostly hyperintense. The time course of the BOLD signal of the area indicated in the preoperative images is showed in C (red line, time course of the task; black line, the variation of the BOLD signal during the task. The red line indicates the GLM model of HRF response and the black one the raw signal registered during the acquisition). In the presurgical examinations (A, D), the cluster of BOLD pixels located inside the lesion (white arrows from A and D to C) correctly correlates with the fMRI paradigm. This cluster, labelled as ‘intralesional’, is no longer evident in the postoperative fMRI study due to the removal of the dysplastic area. The frontal BOLD correlated functional clusters, located in the normal brain bordering the anterior aspect of the lesion in pre and postsurgical examinations, are labelled as ‘perilesional’. Right and left sides are indicated in yellow.

https://doi.org/10.1371/journal.pone.0238389.g001

Structural and functional MRI findings are summarized in Table 1.

Among the 29 patients who underwent fMRI study, twenty-seven patients (93%) performed the WG task, 23 (79%) the RG task and 23 (79%) the comprehension task. Twenty-one (72%) patients underwent all three tasks.

The average Global LI was 0.47 (SD 0.48) for WG task, 0.56 (SD 0.33) for RG task and 0.33 (SD 0.56) for comprehension task. The distribution of atypical lateralization (LI < 0.20) and crossed dominance (dissociation between phonological and syntactic-semantic aspects of language) across the different tasks is summarized in Table 1.

According to the exploratory analysis (Figs 2–4 and S1 Table), we selected as possible predictors for the analysis of associations the variables significantly correlated with the first two dimensions of the FAMD, as follows: a) clinical variables—interictal and ictal scalp EEG, topography of the lesion and the SOZ (side and localization) and their relationship, histology, handedness, epilepsy duration, age at seizure onset, seizure frequency, focal to bilateral tonic-clonic seizures; b) fMRI findings—LIs and intra- and perilesional activations during all language tasks; c) neuropsychological findings—global cognitive abilities, fluency, language and backward and forward span and attention scores.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Clinical variables.

Correlation between quantitative and qualitative variables and the principal dimensions (A) and contribution of the variables to dimensions 1 (B) and 2 (C). The red dashed line on the graph indicates the expected average values, if the contributions were uniform. ^aEpiDur: Epilepsy duration; ExT: extratemporal; Foc: focal; FtoB: focal to bilateral tonic-clonic; Handed: handedness; IctEEG: Ictal EEG, InterEEG: interictal EEG; Le: lesion; SOZ: seizure onset zone; SzOnAge: age at seizure onset; TLE: temporal lobe epilepsy; VS: versus.

https://doi.org/10.1371/journal.pone.0238389.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Functional MRI variables.

Correlation between quantitative and qualitative variables and the principal dimensions (A) and contribution of the variables to dimensions 1 (B) and 2 (C). The red dashed line on the graph indicates the expected average value, if the contributions were uniform. ^a IntraActComp: intralesional activation during comprehension task; IntraActRG: intralesional activation during RG task; IntraActWG: intralesional activation during WG task; LangAct: fMRI activation during all language tasks; NoPeriActRG: absence of perilesional fMRI activation during RG task; NoPeriActWG: absence of perilesional fMRI activation during WG task; PeriActComp: perilesional fMRI activation during comprehension task.

https://doi.org/10.1371/journal.pone.0238389.g003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Neuropsychological variables.

Correlation between quantitative and qualitative variables and the principal dimensions (A) and contribution of the variables to dimensions 1 (B) and 2 (C). The red dashed line on the graph indicates the expected average value, if the contributions were uniform. ^aAttent: attention scores; Langu: language scores: PreNorCogn: preoperative normal cognitive scores; SPANBW: backward span scores; SPANFW: forward span scores.

https://doi.org/10.1371/journal.pone.0238389.g004

We then assessed through univariate analysis (see S2 Table), according to our causal model, whether any of the selected clinical variables may represent a predictor of the fMRI results, thereby disclosing the following significant associations:

• Patients with diffuse or bilateral ictal scalp EEG had a 100% probability of exhibiting an intralesional activation during all language tasks, while patients with focal ipsilateral ictal EEG had 50% probability (Fisher’s exact test, p = 0.012).
• Patients with diffuse or bilateral ictal scalp EEG had a higher probability of disclosing an intralesional or perilesional activation during WG task than patients with focal ipsilateral ictal scalp EEG (OR = 12.83, p = 0.031, LR χ²₍₁₎ = 6.51 p = 0.011, Pseudo R² = 0.19 and OR = 6.27, p = 0.05, LR χ²₍₁₎ = 4.47 p = 0.034, Pseudo R² = 0.09, respectively).
• Patients with intralesional activation during RG task had a higher average duration of epilepsy (OR = 1.25, p = 0.049, LR χ²₍₁₎ = 5.65 p = 0.017, Pseudo R² = 0.17).

The multivariate linear regression model revealed a significant effect of diffuse interictal scalp EEG abnormalities (regr. coeff = -0.401 p = 0.014) and handedness (regr. coeff = -0.768 p = 0.002) on Lis (Table 2). Specifically, language lateralization was more atypical if EEG abnormalities were bilateral or diffuse rather than focal or unilateral and if patients were left-handed rather than right-handed.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Multivariate linear regression model of LIs.

https://doi.org/10.1371/journal.pone.0238389.t002

The likelihood ratio test was significant (LR χ²₍₂₎ = 20.270 p<0.001), thus indicating that our model is more effective than the null model. The proportion of the total variability explained by the model as calculated by the adjusted R² was 48.9%. The sensitivity analysis confirmed the results of the multivariate regression model.

Regression models showed that LI during comprehension task became more atypical when age at seizure onset increased (regr. coeff = -0.051; p = 0.036, LR χ²₍₁₎ = 4.93 p = 0.026, Adj R² = 0.15) and if patients were left-handed (regr. coeff = -0.6; p = 0.020, LR χ²₍₁₎ = 6.02 p = 0.014, Adj R² = 0.19).

Conversely, side and localization of the lesion and of the SOZ, histology, epilepsy duration, age at surgery and seizure frequency did not predict fMRI results.

As a further test included in our causal model, we evaluated, through univariate analysis, whether the pattern of language representation had an impact on cognitive performances, thereby disclosing the following significant associations:

• Patients with intralesional activation during all language tasks had a higher probability of exhibiting a pathological backward span (Fisher’s exact test, p = 0.023).
• Patients with severe cognitive impairment disclosed a more atypical language lateralization (p = 0.0136) compared to those with higher cognitive scores.
• Patients with lower attention scores had a more atypical language lateralization during WG and RG (F = 21.23, p<0.001 and F = 5.10 p = 0.036, respectively) and comprehension tasks (F = 4.73, p = 0.043) compared to those with higher attention scores.

Univariate and multivariate models did not show any significant result.

DTI.

The mean Volume LI for the arcuate fasciculus was -0.071 and the mean FA LI was -0.022.

Histogram analysis did not highlight hemispheric differences in the average distribution of FA values extracted from the arcuate fasciculus. The calculation of arcuate fasciculus LIs for volume and histogram-based measures could not identify any strong asymmetric pattern. LI values rarely exceeded 0.2 for left asymmetry (-0.2 for right asymmetry) (S2 Fig)

We found no correlation between fMRI language dominance and arcuate fasciculus LI values in most patients.

FMRI (Tables 3 and S3).

Comparing pre- and postoperative LI values, we found atypical language lateralization in one vs three patients during WG task, in none vs three patients during RG task, in two vs seven patients during comprehension task. The analysis of variations of preoperative vs postoperative LIs (paired Student’s t-test) is summarized in Table 3. All changes determined a more atypical language representation after surgery compared to preoperative measures. In particular, we found significant LI changes for RG (anterior frontal ROI) and comprehension tasks (global and anterior frontal ROIs).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Average (±sd) of preoperative and postoperative LIs for global, temporal and fronto-dorsal ROIs.

P-value of paired Student’s t-test.

https://doi.org/10.1371/journal.pone.0238389.t003

When analyzing possible clinical predictors of postoperative changes of LI through univariate analysis (S2 Table), we found that abnormal postoperative interictal scalp EEG (p = 0.003), Engel class IB-IV outcome (p = 0.015) and a longer disease duration were significantly associated with a more atypical language representation after surgery. This latter was in turn associated with a worsened fluency after surgery compared to preoperative performances.

In addition, in patients with right epileptogenic lesions, atypical language representation was observed more often than in those with left epileptogenic lesions (mean ±std -0.65±0.46 vs -0.14±0.35; p = 0.049).

Multivariate analysis did not disclose significant results.

DTI.

For the arcuate fasciculus, the mean postoperative Volume LI was-0.037 and the mean postoperative FA LI was -0.018.

We identified no significant postoperative FA and volume LI changes.