Comparison of lateralization and spatial patterns of hemodynamic and neuromagnetic language brain mapping

Highlights

• •fMRI, MEG, and task integration lower lateralization divergences in healthy subjects.
• •Both modalities map an overlapping left-perisylvian language core.
• •Yet patterns diverge: left fronto-parietal in fMRI vs left temporal/opercular in MEG.

Abstract

Objective

To compare and integrate hemispheric lateralization and spatial patterns of language mapping derived from functional magnetic resonance imaging (fMRI) and magnetoencephalography (MEG).

Method

Twenty right-handed healthy adults performed three language tasks (verb/noun generation, phonological/semantic fluency, and sentence completion). A unified framework ensured methodological consistency across modalities: (i) identical tasks; (ii) individual mapping of block-level signal changes in fMRI and broadband (4–40 Hz) oscillatory power decreases in MEG; (iii) subject-level spatial extent-based thresholding of intramodality maps and quantification of intermodality overlap. Laterality indices (LI) were compared and combined across tasks and modalities to assess lateralization discordances. Spatial patterns were obtained by averaging individual maps, while permutation statistics evaluated intermodality spatial differences.

Results

Task- and modality-combined LI reduced intermodality lateralization discordance from 45% to 10%. MEG showed twice the LI variability of fMRI. Overlap was concentrated in left fronto-temporal regions, while spatial differences emerged in left fronto-parietal areas (fMRI) and temporal/opercular regions (MEG).

Conclusion

Although core language areas were identified in both modalities, modality- and task-integration may enhance robustness of language mapping in healthy subjects.

Significance

Lateralization discordances and modality-specific spatial patterns, along with MEG’s higher lateralization variability, suggest that multimodal integration could aid in presurgical clinical decision-making.

1. Introduction

A large number of functional magnetic resonance imaging (fMRI) studies have shaped our understanding of the spatial distribution of the language network in the human brain (Price, 2012, Vigneau et al., 2006). However, fMRI is an indirect measure of brain function. It detects the blood-oxygen-level-dependent (BOLD) signal, which reflects changes in blood oxygenation driven by neurovascular coupling (Kaplan et al., 2020). fMRI has a limited temporal resolution (on the second-scale) but offers millimeter-scale spatial resolution (Logothetis, 2008).

Electrophysiological techniques, such as magnetoencephalography (MEG), have refined our understanding of the temporal and spectral dynamics of the language network (Baillet, 2017). MEG provides a direct measure of the oscillatory neural activity with a millisecond-scale temporal resolution. It can depict the temporal sequence of language processing in the brain (Dikker et al., 2020, Salmelin, 2007) and capture the broad range of frequency bands associated with the linguistic information processing across neocortical areas (Benítez-Burraco and Murphy, 2019, Coolen et al., 2020, Goto et al., 2011, Lam et al., 2016, Siegel et al., 2012). Because MEG source imaging solves an ill-posed inverse problem and is affected by neural source depth and orientation, its effective spatial resolution is typically coarser and more variable than that of fMRI (Baillet et al., 2014, Baillet, 2017).

Both fMRI and MEG are used for functional language brain mapping in the presurgical assessment of brain lesions (Agarwal et al., 2019, Black et al., 2017, Bowyer et al., 2020, Duncan et al., 2016, Dym et al., 2011, Papanicolaou et al., 2014, Petrella et al., 2006, Pirmoradi et al., 2010, Szaflarski et al., 2017). Such mapping pursues a twofold goal. First, it aims to assess hemispheric dominance (i.e., language lateralization) as a non-invasive surrogate for the invasive Wada procedure (Papanicolaou et al., 2014). Second, it is used to locate language areas relative to the brain lesion for surgical treatment planning (Petrella et al., 2006) and to predict postoperative language deficits (Papanicolaou et al., 2014, Szaflarski et al., 2017).

Regarding hemispheric lateralization, fMRI-MEG intermodality concordance has been reported as high (e.g., 75–100 %; Pang et al., 2010) or poor (i.e., no intermodality correlation; Billingsley-Marshall et al., 2007). Importantly, more right-sided or bilateral neural activity changes have been reported in MEG (Fujimaki et al., 2009, Grummich et al., 2006, Herfurth et al., 2022, Pang et al., 2010, Singh et al., 2002), with a lateralization being more variable than fMRI (Youssofzadeh & Babajani-Feremi, 2019). However, detailed accounts of comparative fMRI-MEG lateralization profiles are scarce (e.g., Youssofzadeh & Babajani-Feremi, 2019). In addition, while fMRI and MEG have been viewed as insufficient in isolation to fully capture a neurosurgical language lateralization profile (Kamada et al., 2007), the potential of integrating several tasks (as in fMRI language mapping, see, e.g., Gaillard et al., 2004) and both modalities to refine the complex lateralization evaluation has not been fully exploited.

Mapping concordances have been shown between fMRI and MEG, e.g., in the left prefrontal cortex (Wang et al., 2012), temporal lobe (Billingsley-Marshall et al., 2007), frontotemporal areas (Dale et al., 2000, Grummich et al., 2006, Singh et al., 2002, Vartiainen et al., 2011, Youssofzadeh and Babajani-Feremi, 2019), frontoparietal cortices (Liljeström et al., 2009, Singh et al., 2002), and bilateral occipital lobes (Vartiainen et al., 2011). Qualitatively, decreases in neural oscillatory power colocalize with increases in BOLD signal, particularly within alpha/beta (Singh et al., 2002) and beta frequency bands (Dumitrescu et al., 2025b). Concordance in the beta frequency range has also been assessed in specific regions of interest (ROIs) (Herfurth et al., 2022), such as Broca’s area (Pang et al., 2010). Quantitatively, a moderate correlation has been demonstrated between increases in BOLD signal and decreases in oscillatory power in those frequency bands (Conner et al., 2011, Hipp and Siegel, 2015, Kujala et al., 2014).

Critically, discordances have also been reported. Frontal areas were often pointed out as problematic, as they may not fully spatially align (Dale et al., 2000). Frontal areas tend to be more consistently disclosed by fMRI than MEG, while the opposite is described for temporal areas (Billingsley-Marshall et al., 2007, Kamada et al., 2007, Liljeström et al., 2009, Vartiainen et al., 2011). Accordingly, neural activity detected by MEG but not fMRI has been described in the left anterior temporal areas (Fujimaki et al., 2009). Peaks of signal changes in language-related ROIs are divergent, with distances reported of several centimeters between modalities (Billingsley-Marshall et al., 2007) that are higher in temporal than frontal areas (Herfurth et al., 2022) and more marked at the subject than the group level (Liljeström et al., 2009), and in Wernicke’s area (Herfurth et al., 2022).

A few comparative studies have provided combined fMRI-MEG functional language mapping focused on brain lesions in neurosurgical patients (Ellis et al., 2020, Grummich et al., 2006, Kamada et al., 2007). Others have centered their comparison on specific ROIs (Herfurth et al., 2022, Pang et al., 2010) or used fMRI as a spatial constraint to reconstruct changes in MEG oscillatory dynamics (Dale et al., 2000, Fujimaki et al., 2009). Only a small number of studies have actually provided whole-brain, unconstrained comparative functional language mapping of both modalities in healthy subjects (Liljeström et al., 2009, Singh et al., 2002, Vartiainen et al., 2011, Wang et al., 2012, Youssofzadeh and Babajani-Feremi, 2019), while none have, to the best of our knowledge, formally established the fMRI vs. MEG frontotemporal predominance.

In patients with brain lesions, some intermodality discrepancies may arise from (peri-)lesional alterations of the vascular BOLD response (i.e., neurovascular uncoupling; see e.g., Pak et al., 2017) in functional cortex, where MEG can still detect neuromagnetic oscillatory power changes (e.g., Grummich et al., 2006). Apart from these specific cases, differences between MEG and fMRI mapping may be due to the different nature of the signal detected by each modality. For instance, fMRI is thought to mirror the metabolic demands originating from sustained neural activity, usually averaged over prolonged periods of time, while MEG can detect transient synchronized activity within specific and short post-stimulus periods (e.g., Billingsley-Marshall et al., 2007, Liljeström et al., 2009, Vartiainen et al., 2011).

Critically, previous fMRI-MEG comparative studies presented important methodological discrepancies. Tasks often varied across modalities in terms of, e.g., stimuli, stimulus duration, and control baseline (Billingsley-Marshall et al., 2007, Ellis et al., 2020, Fujimaki et al., 2009, Grummich et al., 2006, Herfurth et al., 2022, Wang et al., 2012, Youssofzadeh and Babajani-Feremi, 2019). In other studies, different tasks were used (Kamada et al., 2007), or different subjects were analyzed (Singh et al., 2002). Some tasks are deemed more appropriate to disclose either the dorsal (expressive) or ventral (receptive) streams (Voets et al., 2024) of the dual stream model of language (Hickok & Poeppel, 2007), justifying the need for a choice of a task tailored to the lesion location (Grummich et al., 2006). However, different tasks yield varying success rates dependent on the modality (e.g., Grummich et al., 2006, Kamada et al., 2007), prompting the combination of different tasks and modalities for an accurate depiction of the language network (e.g., Youssofzadeh & Babajani-Feremi, 2019). Thresholding of fMRI maps was previously mostly done with a fixed p-value, ranging from p < 0.05 uncorrected (Kamada et al., 2007) to p < 0.05 (e.g., Billingsley-Marshall et al., 2007, Singh et al., 2002) or p < 0.001 (e.g., Liljeström et al., 2009) corrected for multiple comparisons (family-wise error − FWE, or false discovery rate − FDR). In contrast, MEG thresholds were often data-driven (e.g., Billingsley-Marshall et al., 2007, Ellis et al., 2020, Grummich et al., 2006) or nonexistent (Singh et al., 2002, Vartiainen et al., 2011), which may obviously have led to sensitivity differences. Indeed, when using fMRI fixed thresholds, the disclosure or non-disclosure of an area is dependent on scanning session-dependent factors such as signal-to-noise ratio (Gorgolewski et al., 2012, Gross and Binder, 2014, Voyvodic, 2006). Closely related to this issue, the evaluation of hemispheric lateralization is highly dependent on the thresholding of individual maps (Seghier, 2008).

The existing literature thus leaves a gap in the comparative description of the hemispheric lateralization and spatial profiles derived from fMRI and MEG functional language mapping. This study aimed to fill this gap in healthy adults while using a common analysis framework, thereby alleviating several of the previous biases. We first reevaluated the fMRI-MEG concordance of lateralization profiles along with a detailed comparative quantification in each modality. We introduced a novel lateralization evaluation method that integrated both modalities and several tasks to reduce intermodality discrepancies. Second, whole-brain spatial profiles accounting for individual variability were depicted for each modality and their overlap. Finally, the fMRI-MEG fronto-temporal dichotomy was statistically assessed to confirm intermodality spatial differences. Overall, this study is expected to contribute to refining multimodal fMRI and MEG functional language mapping in healthy subjects, thereby laying the groundwork for future applications in presurgical contexts.

2. Methods

2.1. Subjects

Twenty healthy adult subjects (mean age: 31.2 ± 8.1 years; range: 22.0 – 51.4 years; 11 females, 9 males) participated in this study. All participants were right-handed native French speakers, with an average handedness score of 93.3 ± 8.8 % (range: 77.8 – 100 %) according to the Edinburgh Handedness Inventory (Oldfield, 1971). None of the participants had a history of neurological, psychiatric, or learning disorders. All were free of fMRI/MEG contraindications.

Written informed consent was obtained from all subjects, who received a small monetary compensation for their participation. The Ethics Committee of the CUB – Hôpital Erasme (Université libre de Bruxelles, Brussels, Belgium; EudraCT/CCB: B406201732244; Reference: P2017/272) approved the study.

Group-level data obtained in these subjects have already been published (Coolen et al., 2024, Dumitrescu et al., 2025a, Dumitrescu et al., 2025b).

2.2. Language tasks

Three commonly used covert language tasks were performed in French, first in MEG and then in fMRI to avoid potential magnetization artifacts in MEG recordings. The task order was randomized for each subject but remained identical across modalities. For each task, the structure and timing of stimulus presentation were kept consistent across modalities and subjects. Still, the pool of stimuli was split, with half randomly assigned to each modality per subject, to avoid learning or repetition effects (Billingsley-Marshall 2007).

Image stimuli appeared in color on a white background, while written stimuli appeared in white Times New Roman font on a black background. Stimuli were projected onto a screen at the foot of the fMRI and MEG scanner beds and made visible to the subjects through a mirror, keeping the visual angle below 7°. The display of stimuli was marked by a specific trigger signal, recorded in a separate channel synchronously with the MEG data and logged in an event log synchronized to the pulse sequence of the fMRI scanner.

Before the scanning sessions, subjects were tested with a short training version of all three tasks using different sets of stimuli. During this training version, participants were asked to provide their answers overtly to ensure they correctly understood the tasks, which was confirmed in all participants.

Fig. 1 depicts a schematic representation of the tasks.

Fig. 1.

Experimental paradigm overview. Schematic representation of the covert language tasks applied in French in both MEG and fMRI, maintaining consistent structure and timing across both modalities. The general task block structure (upper part of each section) presents an overview of the tasks’ block sequence and duration. The block structure (lower part) displays the sequence of stimulus presentation within each block of task condition, i.e., verb generation (VG), image naming (IN), phonological fluency (PF), semantic fluency (SF), and sentence completion (SC), along with rest (R) and control (C) periods. Each block begins with a written cue to clarify the condition type: “Verbes” (Verbs), “Noms de” (Nouns), “Mots par” (Words with), “Names of” (Names of), “Complétez” (Complete), or “Regardez” (Look). Examples of stimuli are provided for each task, with their translations in the main text.

2.2.1. Verb generation and image naming task

This task has been described in detail in Dumitrescu et al., 2025a). The combined verb generation and image naming task (VGINT) was designed as a mixed event-related/block paradigm. Subjects were required to provide a covert single-word response upon presentation of color drawings of common objects and animals, either generating a related verb during the verb generation condition (VG) or naming the image during the image naming condition (IN). A total of 320 images were used: 160 images originating from the revisited Snodgrass and Vanderwart's object pictorial set (Rossion & Pourtois, 2004) for VG, and 160 images from the MultiPic set (Duñabeitia et al., 2017) for IN. Eighty VG and IN images were presented per modality. The task structure consisted of 10 blocks of VG, alternating with 10 blocks of IN, with the order of VG and IN blocks determined randomly per subject, totaling 20 blocks. Each block lasted 33 s, leading to a total task time of 10 min. At the start of each block, a 1-second cue indicated the assignment at hand (“Verbs” or “Nouns”), followed by a 2-second fixation cross. Within each block, eight images were presented, each for 1 s, interspersed by fixation crosses (with variable interstimulus intervals of 2–7 s) during which participants had to provide their answer. Each block ended with a final fixation cross (2–6.5 s).

VGINT primarily aimed to probe language production. The alternating blocks of high-level language conditions allowed for narrowing down the additional controlled lexico-semantic associations with the image object required by VG compared to the simpler IN (Bourguignon et al., 2018). Indeed, the VG-IN contrast has been shown to disclose left inferior frontal areas, and, to a lesser extent, posterior temporal areas in fMRI (Dumitrescu et al., 2025a), both of which are part of the semantic control network (Jackson, 2021).

2.2.2. Phonological and semantic fluency task

The combined phonological and semantic fluency task (PFSFT) was designed as a classic block paradigm (for details, see Dumitrescu et al., 2025b). Subjects were required to covertly generate as many single-word responses as possible during the presentation of a stimulus, either words starting with the presented letter(s) during the phonological fluency condition (PF) or nouns belonging to a given semantic category during the semantic fluency condition (SF). A total of 40 stimuli were used. For PF, 10 single letters (e.g., “B”) and 10 two-letter phonemes (e.g., “CH”) were selected while ensuring the existence of a sufficient number of frequent words starting with those letters based on Lexique 3 (New, 2006). For SF, 20 semantic categories (e.g., “Professions”/”Occupations”, “Pays d’Europe”/”European Countries”) were inspired by Marchal and Nicolas (2003)Ten PF and 10 SF stimuli were presented per modality. The task structure consisted of five iterations of a 51-second PF (or SF, dependent on per-subject randomization) block, followed by a 51-second SF (or PF) block, and by a 27-second rest (R) block (total: 15 blocks, duration: 10 min 45 s). At the start of each block, a 1-second cue indicated the assignment at hand (“Words with” or “Names of”) or the rest period (“Rest”), followed by a 2-second fixation cross. Within a task block, two 24-second stimuli were successively presented, each followed by a 3-second fixation cross. R blocks consisted of a simple 25-second fixation cross, which subjects were instructed to gaze at.

PFSFT also primarily explored language production. This task design allowed for the contrast of two types of fluency (PF and SF) with a resting fixation (R), which is still commonly used as a control condition in clinical settings (Voets et al., 2024). We decided to group PF and SF tasks for the analysis as they both strongly activate the dorsal stream of the language network (Dumitrescu et al., 2025b

2.2.3. Sentence completion task

The sentence completion task (SCT) was designed as a classic block paradigm (for details, see Coolen et al., 2024). Subjects were asked to read incomplete sentences and to generate their endings with one or a few words. A total of 72 incomplete, short, nonstandardized sentences were created and presented in three parts. Thirty-six sentences were presented during the sentence completion condition (SC) in each modality. The first part (P1; e.g., “La poule”/“The hen”) contained the subject, composed of a common noun and its determiner. The predicate was partly stated in the second part (P2; e.g, “pond un”/”lays an”) and had to be completed by the participants in the third part (P3), which prompted the completion by a visual clue (“_________.”). Visual control stimuli (C) were also created, consisting of sequences elaborated to resemble the target sentences, composed only of “*” characters and spaces. Pseudosentences were also presented in a first (e.g., “** *****”) and second part (e.g., “**** **”), followed by the same empty third part (“_________.”), which the subjects were instructed to simply look at. The task structure consisted of six iterations of a 63-second SC block, followed by a 33-second C block, resulting in a total paradigm duration of 9 min and 36 s. At the beginning of each block, a 1-second cue (“Complete” or “Look”) followed by a 2-s fixation cross clarified the assignment at hand. Each SC block contained six sentences, while C blocks contained three pseudosentences. A 1-second fixation cross preceded every (pseudo)sentence, and each part of the latter (i.e., P1, P2, P3) was then shown for 3 s.

SCT probed both the ventral and dorsal streams of language (Coolen et al., 2024) and allowed for a contrast between SC and a gibberish control condition (C) typical for this task (Black et al., 2017).

2.3. Data acquisition

2.3.1. MEG acquisition

Subjects first performed the language tasks during MEG recordings (band-pass: 0.1–330 Hz; sampling rate: 1 kHz) in a lightweight magnetically shielded room (Maxshield^TM, MEGIN, Helsinki, Finland; see De Tiège et al., 2008, for details) using a 306-channel whole-scalp-covering neuromagnetometer (Triux^TM, MEGIN, Helsinki, Finland). Subjects were comfortably positioned in the supine position to minimize head movement artifacts and to provide them with experimental conditions similar to those in fMRI.

Four head-tracking coils monitored the subjects’ head position inside the MEG helmet. The locations of the coils and a minimum of 350 head-surface points (on the scalp, nose, and face) with respect to anatomical fiducials were recorded with an electromagnetic tracker (Fastrak, Polhemus, Colchester, VT, USA) before starting the MEG session.

2.3.2. fMRI acquisition

Participants then performed the language tasks during fMRI recordings on the same day, using a 24-channel head and neck coil, in a hybrid 3 T SIGNA^TM PET-MR scanner (GE Healthcare, Milwaukee, Wisconsin, USA). fMRI data consisted of T2*-weighted, single-shot echo-planar-imaging sequences with whole brain coverage (time of repetition = 3000 ms, time of echo (TE) = 35 ms, flip angle = 90°, field of view = 26 cm, in-plane resolution = 2.7 x 2.7 mm, slice thickness = 3 mm, ascending interleaved acquisition, 43 slices), synchronized to the tasks. Four dummy scans were performed prior to data collection to allow the signal to reach a steady state. Subjects were also comfortably positioned in the supine position and provided with ample acoustic protection, including earplugs and earphones.

For anatomical localization, a 3D T1-weighted gradient-echo sequence (time of repetition = 8.2 ms, time of echo = 3.1 ms, flip angle = 12°, field of view = 24 cm, matrix = 240 x 240, isotropic 1 mm³ voxels) of the head was also acquired for each subject.

2.4. Individual statistical maps

2.4.1. fMRI individual statistical maps

fMRI data were preprocessed using default parameters in SPM12 (Wellcome Department of Cognitive Neurology, Institute of Neurology, London, UK). The fMRI data were first realigned and coregistered to the subjects’ anatomical T1-weighted image. They were then normalized to the Montreal Neurological Institute (MNI) template, based on the six-tissue probability map segmentation of the anatomical T1-weighted image, and resliced to isotropic 2 x 2 x 2 mm voxels. Finally, they were smoothed with an isotropic Gaussian kernel (full width at half-maximum of 6 mm).

Unthresholded t-maps were obtained using standard first-level analysis in SPM12. This consisted of building a general linear model (GLM) of the preprocessed data for each task in which the corresponding block conditions (VG and IN in VGINT; PF, SF, and R in PFSFT; SC and C in SCT) were modeled as regressors, constructed as a boxcar function that mirrored the timing of the blocks, and convolved with the canonical hemodynamic response function. Movement parameters obtained during the realignment step were also introduced as covariates of no interest. Contrasts were computed and tailored to the task designs, using different levels of control conditions: VG-IN (high-level vs. high-level language condition) in VGINT, PFSF-R (two high-level language conditions vs. uncontrolled rest) in PFSFT, and SC-C (high-level language condition vs. low-level visual control) in SCT. Please refer to the task description for the underlying rationale.

2.4.2. MEG individual statistical maps

The raw MEG data were preprocessed using signal space separation (SSS; Taulu et al., 2005) with default parameters to reduce external magnetic interference and correct for head movements (MaxFilter v2.2, MEGIN, Helsinki, Finland). Physiological (e.g., cardiac, blinking, eye movements) and electronic artifacts were identified using Independent Component Analysis (ICA), specifically the FastICA algorithm (Hyvärinen & Oja, 2000), applied to standardized MEG signals, which combined magnetometer and gradiometer data. A rank reduction to 30 components was applied, with a hyperbolic tangent nonlinearity. The ICA was performed on time series filtered between 0.5 and 45 Hz (Vigario et al., 2000), followed by a visual inspection of components. On average, 2.2 ± 0.6 artifact-related components were identified per dataset and regressed out from the full-rank data. The preprocessed data were then bandpass filtered within the frequency range of interest (4–40 Hz). This frequency range encompassed that of the strongest power decreases in MEG with the best fMRI match (5–25 Hz), as described by Singh et al. (2002), and of the most prominent task-related power decreases (15–35 Hz) reported by Pang et al. (2010).

For source reconstruction, individual anatomical 3D T1-weighted images were segmented using Freesurfer (Fischl, 2012; Martinos Center for Biomedical Imaging, Massachusetts, USA). MEG and MRI coordinate systems were co-registered using the three anatomical fiducial points for initial estimation and the head-surface points for the manual refinement of the surface co-registration. MEG forward models were computed based on the single-layer boundary element method, as implemented in the MNE-C software suite (Gramfort et al., 2014; Martinos Center for Biomedical Imaging, Massachusetts, USA). These forward models were based on a source grid obtained from a common 5-mm cubic grid containing 16,102 source locations in the MNI template brain, and deformed onto individual MRIs by applying the non-linear spatial deformation algorithm implemented in SPM12. Three orthogonal current dipoles were placed at each grid point. The forward models were then inverted via minimum norm estimation (MNE; Dale & Sereno, 1993), using an in-house implementation described by Wens et al. (2015). The sensor-space noise covariance matrix was estimated from 5 min of artifact-free data recorded from an empty room preprocessed using SSS and filtered within the 4–40 Hz range. Block-specific data covariance matrices were computed for each experimental condition (VG and IN in VGINT; PF, SF, and R in PFSFT; SC and C in SCT) using concatenated time points for each block type. The MNE regularization parameter was set for each block type according to the signal-to-noise level estimated from these covariances (Wens et al., 2015), and the depth bias was corrected using noise standardization (sLORETA; Pascual-Marqui, 2002). The final MNE inverse operator combined both magnetometer and gradiometer data, resulting in source-projected dipole current data.

The instantaneous amplitude of dipole current time courses was estimated as the Hilbert envelope of the broadband (4–40 Hz) source-projected signals. The amplitude envelopes, which capture relatively slow fluctuations (below 18 Hz, see Cordier et al., 2024), were then downsampled from 1000 Hz to 100 Hz to reduce data size while preserving amplitude dynamics. Block-level amplitude for each source was calculated by averaging each source envelope across all time points within each block, resulting in source-space maps of average amplitude for each block (10 VG and 10 IN blocks in VGINT; 5 PF, 5 SF, and 5 R blocks in PFSFT; 6 SC and 6 C blocks in SCT). This block analysis, although unconventional because it prevents benefiting from the temporal resolution characteristic of MEG, was employed here to obtain a design as close as possible to that of fMRI (as in Singh et al., 2002, Dumitrescu et al., 2025b) while preserving the spectral specificity of MEG signals. The underlying assumption is that band-limited amplitude fluctuations are closely related to BOLD signal changes (Logothetis et al., 2001, Mantini et al., 2007, Sockeel et al., 2016). It also presented the advantage of being free of assumptions regarding specific time–frequency windows of interest.

Unthresholded t-value maps were created using a mass-univariate, two-sided, two-sample t test (ttest2, Matlab 2024a, MathWorks, Natick, MA, USA) using blocks as samples: VG-IN in VGINT, PFSF-R in PFSFT, and SC-C in SCT. These block-level t-tests closely match the GLM design used in fMRI. When comparing both PF and SF against R, the 10 blocks (five PF and five SF blocks) were combined and compared to the five R blocks. Finally, 3D maps with a resolution matching fMRI (2 x 2 x 2 mm) were generated by aligning each of the 16,102 source locations with their corresponding vertices on the MNI template source grid.

2.5. Spatial extent-based normalization of individual maps

Comparing hemodynamic and neuromagnetic language mapping implies matching functional maps derived from distinct physiological processes (i.e., indirect neurovascular coupling in fMRI vs. direct neural oscillatory activity in MEG), different signal properties (T2*-weighted BOLD signal vs. neuromagnetic fields), temporal and spatial resolutions (second and millimeter scales in fMRI vs. millisecond and 5–10 mm scales in MEG), and analysis methodologies (e.g., Baillet et al., 2014, Friston, 2005, Poldrack et al., 2008). In this context, we elected to normalize the individual maps in both modalities by using spatial extent-based thresholding. In fMRI, this thresholding method has been demonstrated to enhance reliability in language mapping (Wilson et al., 2017) and improve the prediction of postoperative language deficits when considering the 10 % most robust voxels (You et al., 2019). It alleviates some of the issues associated with standard fixed thresholds based on t-values, which have been criticized in fMRI (Gorgolewski et al., 2012, Gross and Binder, 2014, Voyvodic, 2006). Indeed, t-values are influenced by several factors, including the scanner-related signal-to-noise ratio of the data (Gorgolewski et al., 2012, Voyvodic, 2006), as well as subject-related factors such as movement and cognitive variability (Gorgolewski et al., 2013). Our approach aimed to allow for proper spatial comparison between the two modalities by focusing on the end result of individual mapping, i.e., maps with an equal amount of the most robust changes.

Spatial extent-based normalization involved first applying a mask that contained only supratentorial cortical parcels from the Automated Anatomical Labeling (AAL) atlas (Tzourio-Mazoyer et al., 2002), referred to hereafter as the AAL mask. This approach excluded subcortical structures, to which MEG is relatively insensitive (Baillet et al., 2014), unlike fMRI, setting a fair spatial canvas. Given the inverse relationship between BOLD response and MEG power (Conner et al., 2011, Herfurth et al., 2022, Kujala et al., 2014, Pang et al., 2010, Singh et al., 2002), positive values (task-correlated BOLD changes) were considered for fMRI t-maps, while negative values (i.e., power decreases) were used for MEG t-maps. These t-maps were converted to spatial extent-based normalized maps using a ranking procedure similar to that of You et al. (2019): voxels with absolute t-values in the top 10 % across all voxels within the AAL mask were assigned a value of 1, while the remaining voxels were assigned a value of 0. If the target percentage could not be achieved, t-maps were binarized, with non-zero values occupying, e.g., 8.86 % of voxels within the AAL mask. This occurred in only 2 out of 120 maps (20 subjects × 3 tasks × 2 modalities) and was considered negligible. Based on our language mapping experience, the chosen spatial extent-based threshold of 10 % was deemed rather lenient, in line with experts’ recommendations to prioritize sensitivity over specificity in language mapping (Voets et al., 2024). An illustration of how lenient the 10 % threshold appears compared to other spatial extent-based thresholds (e.g., 1 %, 5 %, 15 %) is shown in Supplementary Materials − Fig. 1.

The intermodality overlap was computed for each subject and task as the intersection of non-zero voxels between fMRI and MEG normalized binarized maps. We elected to conservatively evaluate the intermodality overlap at the subject level due to the known interindividual variability in the location of language-related areas (see e.g., Ojemann, 1983, Xiong et al., 2000). This avoided the potential overestimation of intermodality overlap of group-level conjunction analyses, which would not necessarily account for the actual amount and location of overlap at the subject level.

2.6. Lateralization profiles

At the individual level, a laterality index (LI) was computed for each subject and spatial extent-based normalized map using the classic formula (L-R)/(L + R) (Seghier, 2008), where L and R are the sum of voxels in left (L) and right (R) hemispheric ROIs. Specific predefined language ROIs were not used due to the documented spatial differences between fMRI and MEG mapping (e.g., Kamada et al., 2007, Liljeström et al., 2009) and to avoid the use of post hoc ROIs based on the results (e.g., Youssofzadeh & Babajani-Feremi, 2019). The laterality cutoff chosen was 0, with positive values indicating left lateralization and negative values right lateralization. In our right-handed subjects, lateralization was considered typical if it was leftward (Tzourio-Mazoyer et al., 2017) and atypical if it was rightward. A novel LI type, based on the fMRI/MEG overlap maps, was calculated to evaluate lateralization based on regions disclosed by both modalities. Moreover, since a combination of tasks is recommended for the robust evaluation of lateralization (Gaillard et al., 2004, Voets et al., 2024), LI was also calculated on the sum of each task’s spatial extent-based normalized maps and intermodality overlap maps to help resolve discrepancies. The concordances and discordances were then evaluated across and within modalities and tasks by comparing the sign of LI.

At the group level, the lateralization strength and consistency of the novel LI based on intermodality overlap maps were compared to the LI derived solely from fMRI and MEG by using repeated-measures analyses of variance (rmANOVA) as implemented in MATLAB 2024a (MathWorks, Natick, MA, USA). The within-subjects factors for the rmANOVA were Map Type (fMRI, MEG, and intermodality overlap maps) and Contrast (VGINT, PFSFT, SCT, and combination of tasks). While the statistical model included the main effects of the factors and their interactions, the factor of interest was Map Type. Descriptive statistics consisted of marginal means and standard deviations of LI values derived from fMRI, MEG, and intermodality overlap maps, adjusted across the contrasts. Post hoc pairwise comparisons were then conducted to quantify differences. To account for potential violations of sphericity, p-values for the rmANOVA were corrected using the Greenhouse-Geisser method (p_GG). Post hoc pairwise comparisons were corrected for multiple comparisons using the Bonferroni method (p_Bonf).

2.7. Intra- and intermodality spatial patterns

As standard group analyses (e.g., SPM second-level analyses) aim to maximize group-level sensitivity by minimizing inter-individual variation (Voets et al., 2024), group effects may not be representative of a given subset or even of any of the individual maps composing that group (Seghier & Price, 2016). For instance, a significant group effect may be driven by a small effect at the individual level, albeit one that is consistently present across all subjects; alternatively, a nonsignificant group effect may be due to an opposing effect in different subgroups (Seghier & Price, 2016). Circumventing this issue, group-level maps were estimated as the mean of the 20 spatial extent-based normalized maps. Therefore, these maps represent the proportion of subjects displaying their top 10 % changes at each voxel, akin to the spatial overlap maps presented by Seghier & Price (2016).

2.8. Spatial differences between fMRI and MEG language mapping

We sought to formally establish the previously reported fMRI-MEG fronto-temporal mapping dichotomy (Billingsley-Marshall et al., 2007, Fujimaki et al., 2009, Kamada et al., 2007, Liljeström et al., 2009, Vartiainen et al., 2011) and locate regions of intermodality divergence. To assess whether fMRI and MEG systematically differed across subjects, we tested for the consistency of intermodality differences at each voxel. Specifically, we aimed to identify voxels where one modality predominated over the other in a significant proportion of subjects.

To this end, we computed individual voxel-wise difference maps of spatial extent-based normalized maps (fMRI minus MEG), assigning a value of + 1 where only fMRI showed changes, −1 where only MEG did, and 0 where both or neither did.

Under the null hypothesis that there is no consistent direction of intermodality difference across subjects, the values (−1,0,+1) of the individual difference maps at each voxel are assumed to be symmetrically distributed around zero. We therefore applied a custom-coded one-sample, two-tailed sign-flipping permutation test with maximum-statistic FWE control (Nichols & Holmes, 2002). For 10,000 permutations, each subject’s entire difference map was multiplied by + 1 or −1, voxel‑wise means were recomputed, and the maximum absolute value across all voxels was recorded to build a family‑wise null distribution. We determined the significance threshold as its 95th percentile and the FWE‑corrected p-values as the proportion of permutations yielding a maximum absolute mean larger than the unpermuted observed mean. The FWE mask was applied to the mean difference maps to highlight voxels where a significantly greater proportion of subjects showed fMRI‑ or MEG‑predominant responses.

2.9. Map overlays

Maps of interest were projected onto a standard MNI brain that had posterior fossa elements removed, using MRIcroGL (Rorden & Brett, 2000).

3. Results

3.1. Comparative fMRI-MEG lateralization profiles

Fig. 2 summarizes the subjects’ hemispheric lateralization profiles. fMRI-MEG concordance (same LI sign) was observed across tasks in 11 out of 20 (55 %) subjects, while intermodality discordance (opposite signs) in at least one task occurred in 9 subjects (45 %). In fMRI, discordance across tasks and atypical rightward lateralization in at least one task were observed in 4 subjects (20 %), whereas this was observed in 7 subjects (35 %) in MEG. When combining fMRI and MEG, discordance across tasks was observed in 4 subjects (20 %). Notably, all subjects were deemed typical and left-lateralized using fMRI and SCT.

Fig. 2.

Comparative fMRI and MEG individual lateralization profiles. For each subject (rows) horizontal color bars represent the laterality indices (LI) expressed in percentages for fMRI (red bars), MEG (blue bars), and the intermodality overlap (fMRI&MEG; green bars), calculated for each task (first three columns) and the combination of tasks (fourth column).

When LI was calculated for the combination of all three tasks, intermodality concordance was observed in 18 subjects (90 %), and discordance in 2 subjects (10 %). Nineteen (95 %) subjects presented typical leftward lateralization using fMRI and fMRI-MEG overlap maps, while 17 subjects did so using MEG maps (85 %). Only the first subject was deemed atypical with rightward lateralization and concordant evaluation from fMRI, MEG, and fMRI-MEG overlap maps.

Group analysis with rmANOVA demonstrated a significant main effect of Map Type on LI: F(2, 38) = 17.98, p_GG < 0.001. LI marginal means (± standard deviation) were higher for fMRI-MEG overlap maps (57.3 ± 5.8 %) than for MEG (35.6 ± 6.0 %) and fMRI (26.5 ± 2.8 %), with standard deviations being approximately twice as high in fMRI-MEG and MEG compared to fMRI. Pairwise comparisons confirmed that fMRI-MEG overlap maps were overall more left-lateralized (p_Bonf < 0.05) than MEG (difference: 21.6 ± 3,2%) and fMRI (difference: 30.8 ± 5.4 %). However, fMRI and MEG did not significantly differ from each other in terms of lateralization (p_Bonf = 0.555).

3.2. Examples of combined fMRI-MEG mapping

Based on the laterality profile results, we sought to illustrate examples of discordant and concordant laterality profiles and to locate the associated changes in five subjects (Fig. 3). Supplementary Materials − Fig. 2 depicts the combined fMRI-MEG mapping of all subjects.

Fig. 3.

Examples of fMRI-MEG multimodal language mapping. Multimodal individual language mapping in five subjects (rows), displaying fMRI (red) and MEG (blue) changes, and their overlap (green), along with their laterality indices in the corresponding colors, using the three tasks (columns). Maps are rendered on external views of the left and right hemispheres for each subject.

Subject 1 presented discordant laterality profiles between modalities and tasks, primarily characterized by clear MEG rightward changes in VGINT in temporo-parietal regions (intersecting with some fMRI parietal changes), and in disagreement with fMRI-MEG concordant left lateralization in fronto-temporal areas in SCT.

Subject 2 displayed concordant left lateralization across modalities and tasks, except in fMRI and PFSFT in fronto-temporo-occipito-parietal regions, where LI was close to zero. Of note, fMRI-MEG overlap remained clearly left-lateralized in a small temporal area in PFSFT. Leftward concordance was also observed for subject 9, with the exception of VGINT in MEG in right fronto-temporo-parietal areas. Here also, the small frontal overlap was clearly left-lateralized.

Subjects 11 and 13 displayed concordant and clear left lateralization across modalities and tasks, with varying overlap in fronto-temporo-parietal regions.

3.3. Comparative fMRI-MEG spatial mapping

Overall, all changes were clearly left-lateralized across modalities and tasks. However, spatial patterns partly differed between fMRI and MEG, with limited intermodality overlap, together with small areas where mapping differed significantly between modalities, as well as some task-specificity. These findings are described hereafter and depicted in Fig. 4 and Supplementary Materials − Fig. 3.

Fig. 4.

Group-level fMRI-MEG comparative language mapping. Group maps representing the mean of individual spatially normalized maps displaying changes in fMRI (first row), MEG (second row), and in both modalities (fMRI&MEG; third row). These maps, thresholded in the range of 10–90% and color-coded from blue to red, represent the proportion of subjects (n = 20) displaying changes. The bottom row (fMRIvsMEG) represents a statistical mask (FWE p < 0.05) indicating areas more consistently mapped by fMRI (in red) or MEG (in blue). The task-specific patterns are displayed in the three columns. Maps are rendered on external views of the left and right hemispheres.

Mean fMRI changes were mainly frontal, and to a lesser extent, parieto-temporal. Frontal changes concerned the supplementary motor area (SMA), the inferior frontal gyrus (IFG), the mid and inferior parts of the precentral gyrus, and the dorsolateral prefrontal cortex, including the middle and superior frontal gyri. Parietal changes primarily involved the inferior parietal lobule (IPL) and were less prominent in SCT. Temporal changes were centered on the posterior middle temporal gyrus (MTG) for VGINT and SCT, while they were centered on the posterior inferior temporal gyrus (ITG) for PFSFT. Additionally, bilateral changes in the occipital cortices were mainly observed in SCT.

Mean MEG changes were mainly temporal and centered on the operculum of the sylvian fissure. Temporal changes were widespread and concerned a large part of the superior temporal gyrus (STG), MTG, and ITG, as well as the adjacent fusiform gyrus (FUS). Sylvian opercular changes involved the IFG, rolandic operculum, and STG, and were more pronounced in SCT, while they were limited in VGINT. Some occipital changes were also seen, mainly in PFSFT.

Mean intermodality overlap was much smaller and rather inconsistent across tasks. Frontal overlap involved parts of the fMRI frontal changes, with a clear predominance in SCT in the IFG and rolandic operculum. Temporal overlap was mainly posterior and predominant in the STG/MTG for VGINT and SCT. Some occipital overlap was mainly seen in PFSFT.

When thresholded for FWE (p < 0.05), fMRI more consistently identified small left‑hemisphere clusters in the IFG in all tasks, as well as in inferior/mid precentral gyrus and SMA for PFSFT and SCT, and in IPL for VGINT and PFSFT. More fMRI changes were also found in bilateral occipital cortices in SCT. In contrast, MEG showed more consistent mapping only in SCT, limited to the rolandic/temporal operculum and mid‑portions of MTG/ITG.

4. Discussion

This study fills a gap in the depiction of lateralization and spatial profiles of combined fMRI-MEG functional language mapping in healthy adults, utilizing a unified methodological framework designed for a balanced intermodality comparison of three well-documented language tasks. It provided: (1) a comparative account of lateralization profiles, introducing a novel method for evaluating lateralization across imaging modalities and tasks that decreases intermodality discordance, and highlighting greater variability in MEG’s lateralization profiles than fMRI’s; (2) a delineation of intermodality overlaps within core left perisylvian language areas; and (3) a depiction of intermodality differences, establishing a fMRI-MEG fronto-temporal gradient.

4.1. Lateralization profiles

Compelling comparative lateralization profiles have been reported by Youssofzadeh & Babajani-Feremi (2019), albeit in only two tasks and while comparing fMRI GLM-derived maps with MEG connectivity maps. This study presents a new account of fMRI-MEG comparative hemispheric lateralization profiles across three commonly used language tasks in healthy adult subjects. As in previous studies, intermodality discordances (up to 45 %) were observed within tasks (e.g., Wang et al., 2012, Youssofzadeh and Babajani-Feremi, 2019) and across tasks (Kamada et al., 2007). However, this study introduces a novel method for integrating multitask and multimodal fMRI-MEG data. Calculating LI on the combined maps of all three tasks decreased intermodality discordance (down to 10 %) and increased the rate of expected leftward lateralization in our right-handed healthy adult subjects (Tzourio-Mazoyer et al., 2017). This aligns with the recommendation to use a panel of language tasks (Voets et al., 2024), which may help resolve discordant hemispheric lateralization profiles in fMRI (Gaillard et al., 2004). Moreover, LIs derived from intermodality overlap maps were more left-lateralizing than those based on fMRI or MEG alone. Therefore, integrating tasks and modalities may be helpful in more confidently assessing language lateralization. This said, while concordance of language mapping with the invasive Wada test is high in both fMRI (83.5 %, as reported in a meta-analysis by Dym et al., 2011) and MEG (71–100 %, as reviewed by Pirmoradi et al., 2010), it remains imperfect and is complicated by the absence of “ground truth” of functional language mapping (Papanicolaou et al., 2014, Voets et al., 2024). Besides, hemispheric language lateralization may be complex and dissociated, e.g., in the precentral and angular gyri (Seghier et al., 2011) or between frontal and fronto-temporal regions (Kamada et al., 2006). Moreover, as suggested by Kamada et al. (2007), considering fMRI and MEG separately may still be necessary due to their non-overlapping spatial profiles. Therefore, a comprehensive and nuanced assessment of language hemispheric lateralization should also probably include the observation of changes separately in all modalities and tasks.

Similar to Youssofzadeh & Babajani-Feremi (2019), we observed that MEG lateralization profiles were more variable, presenting around twice the standard deviation of fMRI. This finding aligns with previous studies, which have reported mostly discordant cases in MEG, where more bilateral or right-lateralized regions were observed (e.g., Grummich et al., 2006, Herfurth et al., 2022, Pang et al., 2010). This was reflected by more right lateralization in MEG (35 % in at least one task) than in fMRI in our study (20 %). Despite averaging the oscillatory activity over blocks, MEG still inherently captures fleeting, synchronized neural activity (Baillet, 2017), whereas fMRI reflects slower, integrated, metabolically driven changes (Logothetis, 2008, Logothetis et al., 2001). Therefore, MEG could be more sensitive to factors such as task performance and cognitive strategy, as oscillatory activity may be more directly linked to these factors than the fMRI BOLD signal (Singh, 2012). Additionally, individual sulcal anatomy variations may affect the ability of MEG to capture relevant neural activity in the sulcal wall or crowns (Baillet, 2017) of language-related regions. Such physiological and anatomical factors may contribute to the higher variability of MEG lateralization profiles, warranting more caution in their interpretation.

4.2. Similarities: Intermodality overlap

The overlap of BOLD increases and power decreases observed were limited but included the left IFG, rolandic operculum, posterior MTG, and ITG, in keeping with core areas of the left perisylvian language network (e.g., Price, 2012) and encompassed both the dorsal and ventral streams of language (Hickok and Poeppel, 2007, Saur et al., 2008). Consistent with previous studies (Herfurth et al., 2022, Pang et al., 2010, Singh et al., 2002), power decreases in MEG reflect the engagement of a cortical area with increased excitability (Pfurtscheller, 2001), as do BOLD task-correlated changes (Logothetis et al., 2001). In the range of frequencies considered in this study (4–40 Hz: theta to low-gamma), a moderate correlation has indeed been established between positive BOLD signal and neural oscillatory power decreases (Conner et al., 2011, Dumitrescu et al., 2025b, Hipp and Siegel, 2015, Kujala et al., 2014). However, this correlation is complex and varies across brain areas and frequency bands (Conner et al., 2011, Hipp and Siegel, 2015, Kujala et al., 2014), but also depends on the tasks (Mononen et al., 2022) and subject (Hipp & Siegel, 2015). The relative paucity of intermodality overlap we observed may partly be attributed to the known limited intermodality correlation, as well as its spatial, spectral, task, and subject-dependent variability.

4.3. Dissimilarities: Intramodality spatial patterns

Clearly, different spatial patterns emerged in fMRI and MEG, aligning with the scarce intermodality overlap of language locations with intraoperative correlation described by Ellis et al. (2020), and supporting the hypothesis of more complementarity than redundancy between fMRI and MEG. In the frontal lobe, the prominent changes in the left SMA, IFG, mid and inferior portion of the precentral gyrus we observed in fMRI contrasted with the smaller changes in the left IFGop and rolandic operculum in MEG. These findings are consistent with the existing literature, which describes more prominent left frontal changes in fMRI than MEG (Billingsley-Marshall et al., 2007, Liljeström et al., 2009, Vartiainen et al., 2011) and misalignment of fMRI-MEG changes in that lobe (Dale et al., 2000, Dumitrescu et al., 2025b). Conversely, in the temporal lobe, the small changes we observed focally in the posterior part of the left MTG/ITG in fMRI contrasted with the widespread changes over the whole STG, MTG, ITG, and FUS in MEG. This aligns with the previously reported supracentimetric distances between MEG and fMRI (Herfurth et al., 2022), marked at the subject level (Liljeström et al., 2009), and the notion of anterior temporal changes detected by MEG but not fMRI (Fujimaki et al., 2009). Our study established that the fMRI-MEG fronto-temporal gradient previously described was statistically significant. More consistent changes detected by fMRI were located in the left IFG, inferior/mid precentral gyrus, and SMA. Conversely, more consistent MEG changes were centered on the fronto-temporal operculum and mid portions of the MTG/ITG, consistent with covert word production (Dimitriadis, 2025, Dumitrescu et al., 2025b). It is worth noting, however, that both intermodality similarities and intramodality dissimilarities exhibit a task-dependent pattern, consistent with the well-documented task-dependent spatial patterns in language mapping (see, e.g., Binder et al., 2008, Black et al., 2017, Voets et al., 2024, Wilson et al., 2017). Specifically, the more prominent left temporal changes in MEG were only disclosed in SCT, which probably probed the receptive aspect/ventral stream of language more than VGINT and PFSFT.

5. Limitations

The tasks used were all visually presented, which limited the analysis to the regions underlying the processing of visual stimuli, e.g., bilateral occipital cortices, as opposed to those elicited by phonological stimuli, e.g., bilateral STG (Price, 2012). Although the tasks were straightforwardly understood by all subjects, as documented during the training session, the covert nature of the tasks did not allow for verification of their correct execution.

Spatial extent-based thresholding is not substantiated by a neurophysiological or formal statistical approach, given the absence of ground truth for pre/intraoperative language mapping (Papanicolaou et al., 2014, Voets et al., 2024) that precludes the use of any reliable gold standard, as well as known concerns with fixed statistical thresholds (Gorgolewski et al., 2012, 2013). Our spatial extent-based approach instead contributed to a balanced intermodality comparison by producing maps of the most robust changes in equal quantities in both modalities. However, it was still only partially fair given that the spatial resolution inherently differs by an order of magnitude between modalities (Baillet et al., 2014, Logothetis, 2008). Rather than representing the true locations and limits of neural activity, the spatial extent and boundaries of fMRI/MEG changes are intrinsic to the technique and post-processing. In fMRI, this includes the spatial imprecision of the scanner, the “smearing” of the vascular BOLD response around the active patch of cortex, and the degree of smoothing of the raw images (Logothetis, 2008). In MEG, even highly focal sources result in spatially blurred maps due to the spatial leakage associated with the ill-posed nature of the inverse solution (Wens, 2015, Wens et al., 2015). Moreover, determining a spatial extent threshold is not trivial (Binder et al., 2008, Wilson et al., 2017), especially when compounded by the inherent uncertainties of the two modalities. The choice of 10 % used in this study was thus arbitrary.

The limited intermodality overlap we observed may be due to more factors than those discussed above. fMRI is thought to be relatively insensitive to transient neural changes due to its reliance on sustained hemodynamic responses, contrary to MEG (Fujimaki et al., 2009, Grummich et al., 2006, Liljeström et al., 2009, Singh et al., 2002). In our study design, the same held true for MEG maps, as power changes were studied at the block level and may have overlooked transient but functionally relevant power changes. Moreover, above the frequency range of this study (4–40 Hz), the gamma band (∼60 Hz) appears as a major contributor to the BOLD signal, as demonstrated by invasive measurements (Conner et al., 2011, Logothetis et al., 2001). The gamma band is considered the most consistent marker of increased BOLD signal in the left frontotemporal cortex during language production tasks (Flinker et al., 2015) and decreased BOLD signal in the default mode network during cognitively intensive tasks (Ossandon et al., 2011). That said, gamma-band oscillations are difficult to reliably measure externally with MEG, especially in the fronto-temporal areas, due to muscle activity (Hari and Salmelin, 2012, Muthukumaraswamy, 2013). Finally, the evaluation of spatial patterns of language mapping is complicated by the limitations of cognitive subtraction that have been long known in fMRI (Friston et al., 1996). Indeed, task execution entails an interplay between positive and negative task-related changes. For instance, in nodes of the DMN that may be more or less negatively impacted by tasks (Seghier & Price, 2012). After contrasting conditions, this results in positive or negative differences that arise from subtracting positive or negative values. A detailed analysis of each task would be necessary to disentangle the origins of the positive values in fMRI and the negative values in MEG considered here, which is beyond the scope of this study.

The number of healthy adult subjects who participated (n = 20) limits the inferences that could be made in the general population. However, this number is similar or higher compared to many previous studies, e.g., Billingsley-Marshall et al. (2007; n = 15), Ellis et al. (2020; n = 19), Liljeström et al. (2009; n = 15), Herfurth et al. (2022; n patients = 25; n healthy subjects = 10), Dale et al. (2000; n = 4), Fujimaki et al. (2009; n = 10), and Vartiainen et al. (2011; n = 15). Our method, which did not rely on traditional random-effect analyses, also limits inferences about the neurophysiological processes at play. It did, though, allow for the exploration of individual and task-specific differences, as traditional group maps may obfuscate intraindividual variability (Seghier & Price, 2016) and do not account for intermodality differences that are more marked at the individual level (Liljeström et al., 2009). The depiction of such variability, however, is critical from a presurgical mapping perspective (Voets et al., 2024).

Finally, since our sample comprised only healthy, right-handed adults, the expected rate of atypical language lateralization was low (10 %; Tzourio-Mazoyer et al., 2017). Most subjects were indeed deemed left-lateralized across modalities and tasks. Our design, however, limits generalizability to left-handed subjects in whom atypical organization is more prevalent (20 %; Tzourio-Mazoyer et al., 2017). This also applies to patients with left-hemisphere lesions or epilepsy, where both non-invasive and invasive mapping show higher rates of atypical lateralization and complex patterns of inter- and intra-hemispheric reorganization (Hamberger & Cole, 2011). To address this, future studies using a unified multimodal framework analysis should focus on such populations. Future studies should also ideally quantify method performance against “gold standard” methods, such as the Wada test and direct cortical stimulation in patients undergoing neurosurgery (see, e.g., Papanicolaou et al., 2014, for a discussion about “gold standard” methods). A comparison with “gold standards” was not available in this study as it focused on healthy subjects. Alternatively, a combination of functional mapping with navigated transcranial magnetic stimulation could also be interesting to validate the non-invasive functional neuroimaging findings (Fang et al., 2019).

6. Conclusion

This study reevaluated the comparative lateralization and spatial patterns expected from fMRI and MEG language mapping in a set of healthy right-handed adult subjects by using a common methodological framework and a panel of well-documented language tasks. We proposed a novel evaluation of lateralization that integrated multimodal and multitask data, thereby reducing intermodality discordance in healthy adult subjects. Hemispheric lateralization profiles were more variable in MEG than in fMRI and therefore warrant cautious interpretation. Spatial patterns were characterized by limited intermodality overlap and were located in left perisylvian core language areas. Importantly, spatial discordances were also demonstrated, with fMRI predominantly mapping left frontoparietal areas, and MEG more widespread left temporal areas, including the fronto-temporal operculum. Notably, some task specificity was found in both intra- and intermodality spatial patterns. Overall, this study suggests the added value of integrating data across neuroimaging modalities and language tasks for a more comprehensive functional language brain mapping in healthy adult subjects. As such, it thus paves the way for future validation in presurgical patients.

Funding sources

No involvement in the present study.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A. Supplementary data

The following are the Supplementary data to this article: