Quantifying cerebellar signal detectability in MEG and EEG in epilepsy using anatomically informed source modeling

Abstract

The cerebellum is increasingly recognized as a key component of large-scale brain networks implicated in epilepsy, yet its electrophysiological characterization remains limited in non-invasive recordings. This limitation arises from the cerebellum’s depth, complex folding, and unfavorable source orientations, which challenge conventional magnetoencephalography (MEG) and electroencephalography (EEG). Here, we quantitatively characterize cerebellar signal detectability across modalities and sensor configurations using anatomically informed source modeling at the population level.

We analyzed clinical MEG and EEG recordings from a large cohort of patients with epilepsy undergoing presurgical evaluation (n = 54), selected from a larger consecutive clinical population. Cerebellar and cerebral source spaces were constructed using subject-specific anatomical models derived from routine clinical MRI, enabling consistent forward modeling across individuals. The signal-to-noise ratio (SNR) was estimated at individual source locations and summarized at the regional level. In addition to clinical superconducting quantum interference device (SQUID)-MEG and EEG, multiple on-scalp optically pumped magnetometer (OPM) configurations were evaluated through simulations, including layouts matched to clinical sensor geometries and layouts optimized for posterior fossa coverage. The effects of source orientation, source-to-sensor distance, and head size on SNR were systematically investigated.

In routine clinical recordings, cerebellar SNR was consistently lower than superficial cortical reference levels, confirming challenges in detecting cerebellar activity using standard SQUID-MEG and EEG. Reducing source-to-sensor distance by placing OPMs at SQUID-equivalent locations, i.e., projecting SQUID sensor locations to the scalp, did not improve cerebellar SNR, indicating that proximity alone is insufficient for better detectability of deeper sources. In contrast, cerebellar-optimized OPM layouts produced substantial SNR gains in posterior cerebellar regions. The effects of source orientation influenced SNR differences between OPM and EEG (under identical sensor/electrode coverage) but were secondary to depth- and geometry-related constraints. Mediation analysis further demonstrated that relative sensor distance significantly mediated OPM-related advantages in posterior cerebellar regions, particularly in individuals with smaller head sizes.

These findings demonstrate that cerebellar signal detectability is governed primarily by anatomical depth and geometry rather than sensor proximity alone. By combining anatomically informed source modeling with flexible, region-specific sensor layouts, this work provides a principled framework for evaluating and improving MEG and EEG sensitivity to cerebellar activity, with implications extending beyond epilepsy to non-invasive mapping of deep and highly folded brain structures.

1. Introduction

Epilepsy is increasingly recognized as a disorder of large-scale neural networks that extend beyond the cerebral cortex (Spencer, 2002). Among these distributed structures, the cerebellum has attracted growing attention for its potential involvement in seizure-related networks (Elder et al., 2025; Streng, 2023; Streng et al., 2025; Streng and Krook-Magnuson, 2021). Although traditionally viewed as a center for motor coordination, the cerebellum maintains extensive reciprocal connections with cortical and subcortical regions implicated in epilepsy (Kandel and Buzsaki, 1993; Krook-Magnuson et al., 2014; Kros et al., 2017; Mitra and Snider, 1975), suggesting a possible role in seizure propagation, modulation, and disease progression (Blumenfeld et al., 2009; Caveness et al., 1977; Kandel and Buzsaki, 1993; Kros et al., 2017; Mitra and Snider, 1975; Niedermeyer and Uematsu, 1974). Importantly, accumulating neuroimaging evidence indicates that cerebellar involvement in epilepsy is region-specific (Hagemann et al., 2002; Oyegbile et al., 2011; Sone et al., 2022). Large-scale structural MRI study of patients with epilepsy showed that cerebellar volume loss is most pronounced in posterior, predominantly non-motor lobules, such as Crus I/II and lobule VIIB, while the anterior sensorimotor lobe, vermis, and flocculonodular lobe are relatively spared across epilepsy syndromes (Kerestes et al., 2024). These posterior cerebellar regions are strongly connected with association and limbic cortices and have been implicated in cognitive and behavioral comorbidities commonly observed in patients with epilepsy (Elder et al., 2025). Despite this emerging perspective, cerebellar electrophysiology in humans remains poorly characterized.

Investigating cerebellar activity using non-invasive electrophysiology is technically challenging. The cerebellar cortex lies deep within the posterior fossa and exhibits an exceptionally folded geometry, leading to pronounced orientation dispersion and signal cancellation, as well as reduced sensitivity due to large source-to-sensor distances (Samuelsson et al., 2020b). Routine presurgical evaluations using electroencephalography (EEG) and superconducting quantum interference device (SQUID)-based magnetoencephalography (MEG) are primarily optimized for cerebral coverage. Even clinical magnetic resonance imaging (MRI) may incompletely cover inferior cerebellar regions, limiting accurate anatomical modeling and coregistration (Chen et al., 2025).

Signal-to-noise ratio (SNR) is a central determinant of detectability in non-invasive electrophysiological recordings (Samuelsson et al., 2021). Previous studies comparing SNR between SQUID-MEG and EEG have largely focused on the cerebral cortex and have typically enrolled only small numbers of subjects (Ahlfors et al., 2010; Goldenholz et al., 2009; Hill et al., 2020; Hunold et al., 2016; Iivanainen et al., 2017; Marhl et al., 2022; Piastra et al., 2021). Such limited samples cannot adequately capture interindividual variability in head size, anatomy, and sensor positioning, all of which strongly influence measured signals. In particular, source orientation plays a critical role in determining relative sensitivity between MEG and EEG (Hämäläinen et al., 1993), whereas source depth is a dominant factor when comparing on-scalp optically pumped magnetometer (OPM) configurations with conventional SQUID systems. Population-level estimation of SNR across a large clinical cohort, in contrast, provides a more reliable and generalizable understanding of modality performance under realistic clinical conditions.

This population-based perspective is especially important for the cerebellum. Cerebellar geometry and source orientation can vary markedly across individuals, and the depth of cerebellar sources places them at the limit of detectability for conventional systems. In addition, it is known that there is functional asymmetry in the cerebellum that may covary with cerebral asymmetry (Hu et al., 2008; Wang et al., 2013). To date, no systematic, large-scale evaluation of cerebellar SNR across electrophysiological modalities has been established, limiting both the mechanistic understanding of cerebellar signal propagation and measurement at sensor level (based on biophysical principles) as well as its translation to clinical practice.

Recent advances in sensor technology and anatomical modeling provide an opportunity to address this gap. OPMs can be positioned directly on the scalp, eliminating the constrains of traditional cryogenic SQUID helmets and reducing the source-to-sensor distance (Alem et al., 2023; Bagic et al., 2023; Boto et al., 2018; Jas et al., 2021; Pedersen et al., 2022; Sheng et al., 2017). This flexibility enables individualized sensor layouts that may better target deep structures such as the cerebellum, raising the possibility of improved detectability compared with fixed-geometry systems (Lin et al., 2019). However, whether proximity alone is sufficient to overcome depth-related limitations remains unclear.

In parallel, advances in automated neuroanatomical modeling now allow more accurate reconstruction of the cerebellum from routine clinical MRI. Our framework, ARCUS (Samuelsson et al., 2026; Samuelsson et al., 2020a), was developed using a high-resolution ex-vivo cerebellar template to preserve the folding geometry that is difficult to resolve in standard MRI (Sereno et al., 2020). By diffeomorphically registering this template to individual anatomy, ARCUS provides an approximate but highly detailed cerebellar source space that captures the cerebellar cortical folding, making it suitable for electrophysiological forward modeling. When combined with realistic forward and noise modeling, this approach enables systematic estimation of modality-specific SNR in a matter that reflects each patients’ anatomy.

In this study, we apply ARCUS to clinical MRI data from patients with epilepsy to evaluate cerebellar SNR across EEG, SQUID-MEG and simulated OPM-MEG configurations. Cerebral SNR is quantified at the population level to provide a reference for overall modality performance, whereas detailed region-based analyses are focused on the cerebellum. Using simulation-based forward modeling, we assess how sensor modality, geometry, coverage, source orientation, and depth jointly influence detectability, and whether cerebellar-optimized OPM layouts can improve sensitivity relative to conventional systems.

By focusing on modality-specific and region-specific cerebellar SNR and its anatomical determinants, this work establishes a framework for individualized cerebellar electrophysiology in the patients with epilepsy. More broadly, it demonstrates how anatomically informed modeling combined with flexible sensor technology extends non-invasive electrophysiological sensitivity beyond the cerebrum, toward measuring deep brain structures that have traditionally remained difficult to assess in clinical practice.

2. Materials and methods

2.1. Subjects

This retrospective study included patient with intractable epilepsy who underwent routine presurgical evaluation with clinical SQUID-MEG between May 2021 and May 2022 at the Athinoula A. Martinos Center for Biomedical Imaging (Matsubara et al., 2024). All recordings were acquired as part of standard clinical care. This study was approved by Massachusetts General Brigham’s institutional review board (#2010P001169).

2.2. Data acquisition

MEG recordings were obtained using standardized clinical procedure (Tanaka et al., 2024). Participants were positioned supine inside a three-layer magnetically shielded room, with a head support spoon and additional cushioning to maximize comfort and minimize the head motion. To optimize signal quality, the head was carefully positioned at the center of the MEG helmet and seated as deeply as possible within the rigid helmet. Continuous recordings of approximately 50 min were acquired.

When tolerated, a 70-channel EEG cap (EasyCap) was applied for simultaneous EEG recording, together with electrooculogram and electrocardiogram. Some patients declined EEG measurement.

Recordings were performed with a whole-head 306-channel cryogenic SQUID system (VectorView, Elekta Neuromag, Helsinki, Finland), comprising 102 magnetometers and 204 planar gradiometers. EEG electrode locations and additional scalp surface points were digitized using the Polhemus Fastrak (Colchester, VT, USA). For interictal spike localization and anatomical modeling, individual T1-weighted MRI scans acquired at referring institutions were used. Because MRI data were obtained clinically across multiple sites, acquisition parameters and image quality varied.

For the cerebellar reconstruction (Section 2.4), individual MRI scans were first processed with FreeSurfer (version 7.4.1, default parameters) (Dale et al., 1999). Because ARCUS relies on anatomically valid whole-brain surface reconstruction as a prerequisite, initial inclusion and exclusion criteria were defined to ensure sufficient image quality for standard FreeSurfer processing at the cerebral level. Inclusion criteria comprised of: (i) availability of individual T1-weighted MRI; and (ii) isotropic or near-isotropic voxel size < 5 mm. A total of 109 patients initially met these criteria. Exclusion criteria included: (i) gross cerebral or cerebellar abnormalities (e.g., tumor, surgical cavity, cyst); (ii) failure of FreeSurfer reconstruction due to artifacts (e.g., shunt) or poor image quality insufficient for reliable whole-brain surface reconstruction required for clinical MEG analysis; (iii) inability to construct a valid three-layer boundary element (BEM) model due to incomplete head coverage; or (iv) inability to acquire EEG.

After applying these criteria, 62 patients were included for the subsequent ARCUS processing (mean ± SD age = 26.2 ± 14.0 years, range 4.7–61.4 years; 25 males). MRI data were acquired on 3T scanners in 59 patients and on 1.5 T scanners in 3 patients. Slice thickness was < 1 mm in 19 patients, 1–1.2 mm in 31 patients, and ≥ 1.2 mm in 12 patients (range 0.84–4.5 mm). Among excluded patients, exclusion reasons comprised of structural abnormalities (n = 9), FreeSurfer’s failure to reconstruct surfaces or poor image quality (n = 18), BEM construction failure (n = 11), and inability to acquire EEG (n = 10).

To provide context for the following analyses, subject-specific forward models were constructed for MEG and EEG using cortical and cerebellar sources spaces, with cerebellar geometry modeled using the ARCUS framework. SNR was computed at individual source locations across multiple sensor layouts and subsequently summarized at the regional level. Additional analyses examined the effects of source orientation and sensor distance in selected configurations.

2.3. Forward modeling

The MEG and EEG forward models were constructed based on neural current sources located on cortical surfaces. In the cerebrum, dendritic current in pyramidal neurons were modeled as equivalent current dipoles oriented perpendicular to the cortical mantle (Samuelsson et al., 2021). In the cerebellum, which contains approximately 80% of all neurons in the brain, Purkinje cells—characterized by large planar dendritic trees oriented perpendicular to the local cerebellar surface—are considered the dominant contributors to MEG and EEG signals of cerebellar origin (Samuelsson et al., 2020b; Sereno et al., 2020). The same orientation constraint—perpendicular to the local tissue surface—was applied to both cerebrum and cerebellum, enabling direct comparison of SNR across regions within a unified forward modeling framework.

A three-layer BEM model was generated using MNE-Python (Gramfort et al., 2013) with conductivities of 0.3, 0.006, and 0.3 S/m for the brain, skull and scalp, respectively. The cerebral surface was reconstructed using about 100,000 vertices per hemisphere. Because the present source spaces were defined on cortical sheets rather than by gross structure volume, spatial sampling was characterized using surface area and area per vertex (Supplementary Table 1). Across subjects, the median surface area was 182,340 mm² [range, 130,056–227,104] for the cerebrum and 120,451 mm² [81,523–136,752] for the cerebellum, corresponding to a median cerebellum-to-cerebrum surface area ratio of 0.65 [0.51–0.91]. The median area per vertex was 0.82 mm²/vertex [0.69–0.89] for the cerebrum and 0.83 mm²/vertex [0.73–1.23] for the cerebellum. These spatial sampling characteristics indicate that the cerebral and cerebellar source spaces were sampled at a broadly comparable spatial scale.

2.4. ARCUS for cerebellar reconstruction and segmentation

ARCUS (Samuelsson et al., 2026; Samuelsson et al., 2020a) is an automated algorithm for cerebellar segmentation and surface reconstruction (Fig. 1). It employs subcortical segmentation surfaces extracted from FreeSurfer to isolate the cerebellum, followed by non-linear diffeomorphic registration of a high-resolution ex-vivo cerebellar template to the subject’s MRI. This template-driven morphing enables approximate reconstruction of the cerebellar cortical surface even at typical clinical MRI resolutions (1 mm isotropic), which is insufficient for direct folium-level segmentation.

Fig. 1.

ARCUS-derived segmentation and reconstruction.

Representative example of ARCUS-derived cerebellar reconstruction and lobular segmentation from clinical MRI, mapped onto the hig-resolution cerebellar template. Colors denote cerebellar regions and are used consistently across all subsequent figures.

Briefly, the method applies an affine alignment of the ex-vivo template to the subject’s cerebellar volume, followed by non-linear registration using symmetric normalization with cross-correlation (SyNCC, ANTs implementation) (Avants et al., 2008; Avants et al., 2011; Samuelsson et al., 2026). The resulting deformation fields are applied to the template surface to generate a subject-specific cerebellar surface mesh.

In this study, ARCUS was applied within an MNE-Python-compatible workflow using dense spatial sampling, yielding 99,856 vertices (Gramfort et al., 2013).

One of the current study’s objectives was to evaluate the feasibility of applying ARCUS to clinical MRI data of heterogeneous quality. To ensure anatomical plausibility at the spatial scale relevant for this study, ARCUS-derived cerebellar segmentations and source-space representations were visually inspected by an experienced neurophysiologist and neurosurgeon (T.M.). Visual inspection focused on the consistency of lobular boundaries and overall cerebellar morphology with individual MRI, rather than on folium-level surface accuracy.

2.5. Signal-to-noise ratio (SNR) for individual sources

Previous on-scalp MEG studies have primarily characterized sensitivity at the sensor or vertex level, or in terms of field-of-view and information capacity, focusing on the cerebral cortex and typically based on small subject cohorts (Hill et al., 2020; Iivanainen et al., 2017). While these approaches provide important system-level insights, they do not quantify sensitivity at the level of anatomically defined brain regions, nor do they address cerebellar sources in a population-based framework. Here, SNR was evaluated for individual source localization and subsequently summarized within anatomically defined regions, enabling region-wise comparison across modalities and sensor configurations in a large clinical cohort.

SNR was quantified following the formulation of (Goldenholz et al., 2009), extended here to include cerebellar sources and simulated OPM configurations. For each source location, the SNR was defined in decibel (dB) units as:

$$\mathit{SNR}=10{\mathit{log}}_{10}\left[\frac{a^2}{N}\sum _k\frac{b_k^2}{s_k^2}\right],$$

where b_k denotes the signal at sensor k provided by the forward model for a unit-amplitude source, a is the assumed source amplitude, N is the number of sensors, and $s_k^2$ is the noise variance at sensor k. The source amplitude was fixed at 10 nAm and applied uniformly across all sources and modalities.

This formulation yielded an SNR estimate for each vertex across both cerebral and cerebellar source spaces. Vertex-wise SNR reflects the combined effects of source-to-sensor distance, source orientation, and head geometry encoded in the forward model, together with the assumed noise structured described in Section 2.6.

For cross-modality comparisons, SNR difference was defined as the vertex-wise subtraction of SNR values between modalities (in dB), with positive values indicating higher SNR in the first modality.

Since OPM recordings were not available from patients in this cohort, OPM SNR was estimated by simulation using subject-specific forward models generated under multiple sensor-configuration conditions (Section 2.7).

2.6. Noise estimation

Noise variance estimation followed the procedure described by (Goldenholz et al., 2009). Specifically, the noise variance $s_k^2$ for each sensor was estimated from spontaneous clinical MEG and EEG recordings. Sixty seconds of resting-state data were extracted and band-pass filtered between 0.5 Hz and 100 Hz. Five patients were excluded from this analysis and all subsequent analyses because frequent interictal spikes during the resting state recording prevented reliable estimation of baseline noise.

Following the Goldenholz’s framework, the noise variance at sensor k was modeled as:

$$s_k^2=s_k^2{\left({\mathit{AA}}^T\right)}_{\mathit{kk}},$$

where A denotes the forward-solution matrix for unit-amplitude noise sources and $s_k^2$ represents the variance of cortical source activity. This model assumes a spatially uniform distribution of independent noise sources, i.e., equal strength in all vertices of the source-space, oriented perpendicular to the cortical surface, with Gaussian-distributed amplitudes of zero mean and variance $s_s^2$. Instrumentation noise was not explicitly modeled, such that the estimated noise variance reflects biologically plausible background brain activity as mediated by the forward model.

For each modality, median normalized values of ${\widehat{s}}_{k,\mathit{recorded}}^2/{\left(A{A}^T\right)}_{\mathit{kk}}$ were computed separately for SQUID magnetometers, SQUID gradiometers and EEG electrodes. The weighted mean of these three medians, proportional to the number of sensors in each category, was taken as the overall estimate of s_s². This procedure accounts for realistic brain-noise contributions while minimizing bias from individual bad channels. The resulting brain-noise estimate was applied consistently for the computation of SNR in SQUID-MEG, EEG, and simulated OPM configurations. Validation supporting the application of this common noise estimate to OPM simulations is described in Supplementary material.

2.7. OPM configurations

Because OPM recordings were not available from individual patients with epilepsy, multiple simulated OPM configurations were generated to estimate SNR for both the cerebrum and cerebellum, and to enable comparison with clinical MEG (Fig. 2A; SQUID_clinical) and clinical EEG (Fig. 2B; EEG_clinical). For consistency, all simulations used FieldLine Gen1 sensors.

Fig. 2.

Sensor configurations.

(A) SQUID_clinical: Clinical whole-head SQUID system consisting of 306 channels (102 magnetometers and 204 gradiometers). (B) EEG_clinical: Clinical EEG system using 70 scalp electrodes. (C) OPM_SQUID: Simulated OPM layout with sensors positioned at scalp locations corresponding to the clinical SQUID array projected onto the head surface with a 5 mm offset. (D) OPM_EEG: Simulated OPM layout with sensors positioned at scalp locations corresponding to the clinical EEG electrode positions. (E) OPM_cerebellum: Cerebellar-optimized OPM layout consisting of a 5 × 4 grid (10 by 8 cm) placed over the occipital region, with the superior edge aligned to the inion. (F) SQUID_cerebellum: SQUID layout selected for comparison with OPM_cerebellum, consisting of 20 fixed SQUID magnetometer sensors covering the occipital region.

Three OPM configurations were generated:

1. OPM_SQUID (Fig. 2C): To enable a comparison with the clinical SQUID system, 102 virtual OPM sensors were positioned at locations corresponding to SQUID magnetometer positions projected onto the scalp surface, with a fixed 5 mm offset. SNR difference was computed as the subtraction between source-space SNRs for OPM_SQUID and SQUID_clinical.

2. OPM_EEG (Fig. 2D): To compare OPM and EEG under equivalent spatial coverage, 70 OPM sensors were placed at the same scalp locations as standard EEG electrodes based on Polhemus digitization. This configuration enabled direct comparison between OPM and EEG channels using on-scalp sensor placement, with source-space SNR difference between OPM_EEG and EEG_clinical (defined as OPM_EEG minus EEG_clinical SNR).

3. OPM_cerebellum and SQUID_cerebellum (Fig. 2E, F): Whereas the first two configurations reflect conventional clinical layouts optimized for cerebral coverage, this configuration was designed specifically to enhance posterior fossa and cerebellar sensitivity. Twenty OPM sensors were arranged in a 5 × 4 grid (10 cm × 8 cm) over the occipital region. The upper edge of the grid was aligned with the inion, as identified on each subject’s MRI, to anchor the array anatomically relative to the cerebellum. For fair comparison, an equivalent SQUID_cerebellum configuration was defined by selecting twenty fixed SQUID sensors covering the occipital region. Because SNR was normalized by sensor count, the same number of sensors was used for both systems. The SNR difference was computed between OPM_cerebellum and SQUID_cerebellum. Analyses for this configuration were restricted to cerebellar regions.

2.8. Source orientation (folded angle)

To investigate whether regional variability in source orientation contributes to differences in signal detectability between EEG and MEG comparison, we analyzed the relationship between local source orientation and SNR difference in the OPM_EEG-EEG_Clinical comparison scenario. For each subject and vertex, source orientation was quantified as the angle between the surface normal vector and a vector extending from the center of the inner skull surface toward that vertex. Angles were folded into the 0–90° range, where 90° corresponds to a tangential orientation relative to the skull surface (generally favorable for MEG detection), and 0° corresponds to radial orientation.

2.9. Sensor distance ratio

To evaluate how anatomical and geometric factors contribute to SNR differences between OPM_cerebellum and SQUID_cerebellum, we quantified the relative proximity of sensors to cerebellar sources. For each cerebellar vertex, the shortest Euclidean distance to both OPM and SQUID sensors was computed. The sensor distance ratio was defined as the SQUID distance divided by the OPM distance, such that larger values indicate proximity advantage for OPM sensors. Since OPM sensors are positioned closer to the scalp compared to SQUID sensors, the distance ratio was consistently greater than one across all vertices.

2.10. Statistical analysis

2.10.1. Regions of interest

For each metric (SNR, SNR difference, folded angle, and sensor distance ratio), values were averaged within regions of interest (ROIs), combining left and right hemispheres. The sensor distance ratio was only computed for cerebellar ROIs. Regional SNR difference was tested against zero using the Wilcoxon signed-rank test, with false discovery rate (FDR) correction applied separately for cerebrum and cerebellum. Statistical significance was assessed using a two-sided significance level of 0.05 after correction.

For the cerebrum, ROIs were defined to provide a reference framework for cerebellar comparisons rather than a detailed cortical analysis. Accordingly, broad cortical regions were used: frontal, lateral temporal, medial temporal, parietal, occipital, cingulate, and insular regions. These labels were derived from FreeSurfer Desikan-Killiany cortical parcellation. The medial temporal ROI was constructed by combining entorhinal, parahippocampal, fusiform and temporal pole regions.

For the cerebellum, a full set of ARCUS-derived ROIs was analyzed to enable detailed regional evaluation. These included lobules I-III, IV, V, VI, Crus I, Crus II, VIIb, VIIIa, VIIIb, IX, X, as well as vermis VI, vermis VII, vermis VIII, vermis IX, and vermis X.

2.10.2. Distance

To investigate anatomical and geometric contributions to SNR differences between OPM_cerebellum and SQUID_cerebellum, three variables were examined across cerebellar regions: (A) SNR difference, defined in Section 2.7.iii as the regional cerebellar SNR of OPM_cerebellum minus that of SQUID; (B) Sensor distance ratio, defined in Section 2.9.; and (C) Head circumference, used as a proxy for lateral head size.

Head circumference was included as an anatomical variable that constrains sensor-source geometry but does not directly represent sensor placement. In contrast, sensor distance ratio reflects the geometric consequence of both head size and sensor layout on relative sensor proximity.

Pairwise Pearson correlations and linear regression analyses were first performed to characterize relationships among these values. To test whether the effect of the head size on SNR difference operates indirectly through relative sensor distance, mediation analysis (Baron and Kenny, 1986) was performed for each cerebellar ROI. Specifically, we estimated:

1. The a-path, representing the effect of head circumference on the sensor distance ratio.

2. The b-path, representing the effect of the sensor distance ratio on SNR difference after controlling for head circumference; and

3. The indirect effect (a × b), representing the mediated influence of head size through the sensor distance ratio.

Statistical significance of indirect effects was evaluated using Sobel’s test, with FDR correction applied across cerebellar ROIs.

2.10.3. Source orientation

To evaluate whether source orientation contributed to the regional SNR patterns in the OPM-EEG comparison (OPM_EEG vs EEG_Clinical) where sensor geometry was identical, linear regression was applied using all pooled data across subjects and regions. Folded angle was regressed against SNR difference, and statistical significance of the regression slope was assessed using SciPy’s linregress function (Virtanen et al., 2020).

3. Results

3.1. ARCUS feasibility on clinical MRI data

ARCUS processing was performed in 62 patients. Visual inspection confirmed anatomically plausible cerebellar segmentation in 59 of 62 patients (success rate of 95.1%). Representative examples of successful segmentations and all failure cases are shown in Supplementary Figs. 2, and 3, with a summary provided in Table 1.

Table 1:

Examples of successful ARCUS segmentation (Supple Fig. 2) and failed cases (Supple Fig. 3).

Examples of successful ARCUS segmentation
Age (years)	Field strength (T)		Slice thickness (mm)
24	3		1
40	1.5		1
40	3		4.5
Failed ARCUS segmentation
Age (years)	Field strength (T)	Slice thickness (mm)	Finding
43	3	2.2	Partial loss of left VIIb and VIIIa regions
13	3	0.9	Partial loss of right VIIb and VIIIa regions
15	1.5	1.2	Excessively enlarged white matter

Successful segmentations were obtained even in several cases acquired at 1.5 T or with relatively thick slice thicknesses (up to 4.5 mm). In contrast, the failed cases were characterized by poor tissue contrast, particularly reduced distinction between cerebellar cortex and white matter. In one failed 1.5 T case, this blurring likely contributed to loss of cerebellar surface definition.

Overall, the high success rate across heterogeneous clinical MRI datasets supports the feasibility of applying ARCUS to routine clinical imaging for cerebellar source modeling. For subsequent analyses, five patients were excluded due to unreliable noise estimation (Section 2.6.), resulting in a final sample of 54 patients.

3.2. SNR across modalities

3.2.1. SQUID_clinical vs. EEG_clinical

For the cerebrum, the median SNR in SQUID_clinical ranged from −32 dB to −23 dB (Fig. 3A). The highest SNR was observed in the parietal lobe (−23 dB, dashed line), consistent with the superficial location of parietal cortical sources and their favorable orientation relative to the MEG sensors (Laohathai et al., 2021). Although cerebral SNR patterns are well established, they are shown here to provide an internal reference for interpreting cerebellar SNR on the same physical and anatomical scale. In this context, parietal cortex represents a best-case cortical scenario under routine clinical SQUID-MEG acquisition conditions.

Fig. 3.

Cerebral SNR and SNR difference across modalities.

Cerebral SNR maps for (A) SQUID_clinical, (B) EEG_clinical, (C) OPM_SQUID, and (D) OPM_EEG. SNR difference maps comparing (E) SQUID_clinical vs. EEG_clinical, (F) OPM_SQUID vs. SQUID_clinical, and (G) OPM_EEG vs. EEG_clinical.

Asterisks indicate statistically significant SNR differences after false discovery rate (FDR) correction (* < 0.05, ** < 0.01, ***< 0.001). The boxplots show the SNR for each region, in the following order: frontal, lateral temporal, medial temporal, parietal, occipital, cingulate and insula. The dashed horizontal line indicate the median SNR of the parietal cortex in SQUID_clinical, used as a reference for best-case superficial cortical detectability.

Across the cerebrum, regional SNR patterns followed cortical folding geometry, with higher SNR with sulcal walls and lower SNR in deeper structures, including the interhemispheric and Sylvian fissure. In contrast, EEG_clinical showed less regional variability across the cerebrum (Fig. 3B), resulting in a more uniform SNR distribution. The resulting SNR difference indicated that parietal regions—and to a lesser extent frontal and lateral temporal cortices—favored SQUID_clinical, whereas most other cerebral regions—including the cingulate and insular–favored EEG (Fig. 3E), consistent with EEG’s greater sensitivity to deeper sources.

For the cerebellum, median SNR in both SQUID_clinical (Fig. 4A), and EEG_clinical (Fig. 4B) remained below the −23 dB reference level derived from superficial cerebral cortex. While this threshold does not define a detectability threshold in a statistical sense, it provides an empirically grounded benchmark for signals that are routinely detectable in clinical MEG practice. Relative to this reference, cerebellar SNRs were consistently lower across ROIs, indicating the cerebellar signals are recorded under substantially less favorable conditions in standard clinical setups. The SNR difference further showed that Crus I, Crus II, lobules VIIb, VIIIa, VIIIb, IX and X exhibited relatively higher SNR in SQUID_clinical compared with EEG_clinical (Fig. 4E).

Fig. 4.

Cerebellar SNR and SNR difference across modalities.

Cerebellar SNR maps for (A) SQUID_clinical, (B) EEG_clinical, (C) OPM_SQUID, and (D) OPM_EEG. SNR difference maps comparing (E) SQUID_clinical vs. EEG_clinical, (F) OPM_SQUID vs. SQUID_clinical, and (G) OPM_EEG vs. EEG_clinical.

Asterisks indicate statistically significant SNR differences after FDR correction (* < 0.05, ** < 0.01, ***< 0.001). The boxplots show the SNR for each region, in the following order: lobules I-III, IV, V, VI, Crus I, Crus II, VIIb, VIIIa, VIIIb, IX, X, vermis VI, VII, VIII, IX, and X. The dashed horizontal line indicate the median SNR of the parietal cortex in SQUID_clinical, used as a reference for best-case superficial cortical detectability.

3.2.2. OPMS_QUID vs. SQUID_clinical

In OPMS_QUID configuration, OPM sensors were virtually positioned at the scalp locations corresponding to SQUID magnetometer positions projected onto the scalp surface, allowing comparison under the same acquisition modality while reducing source-to-sensor distance. For both the cerebrum and the cerebellum, the SNR maps from OPM_SQUID (Fig. 3C and Fig. 4C) closely resembled those from SQUID_clinical (Fig. 3A and Fig. 4A). The SNR difference showed only minimal differences, with no regions demonstrating higher for OPM_SQUID in either structure (Fig. 3F and Fig. 4F). These results indicate that merely reducing the sensor distance, while preserving geometry and coverage, does not yield meaningful SNR improvement for deeper sources.

3.2.3. OPM_EEG vs. EEG_clinical

In the OPM_EEG configuration, OPM sensors were placed at the same scalp locations as clinical EEG electrodes, enabling direct comparison under equivalent spatial coverage with on-scalp sensor placement. The results (Fig. 3D, G, 4D, G) mirrored those observed in the SQUID_clinical–EEG_clinical (Section 3.2.1): only superficial regions, such as the parietal lobe and Crus I/II, showed higher SNR in OPM_EEG.

3.2.4. OPM_cerebellum vs. SQUID_cerebellum

Fig. 5A–C shows the comparison between OPM_cerebellum and SQUID_cerebellum using layouts specifically optimized for posterior fossa coverage. Since this configuration targets the cerebellum, analyses were restricted to cerebellar regions. Compared with the clinical configuration (Fig. 4A), the OPM_cerebellum layout yielded several posterior regions—Crus I, Crus II, lobules VIIb, VIIIa, VIIIb—with SNR values exceeding the −23 dB superficial cerebral reference level (Fig. 5A). The SNR difference (Fig. 5C) further showed that these regions except for VIIIb, as well as vermis VII and vermis VIII exhibited higher SNR, favoring OPM.

Fig. 5.

Cerebellar-optimized comparison between OPM_cerebellum and SQUID_cerebellum.

(A) SNR for OPM_cerebellum and (B) SNR for SQUID_cerebellum. (C) SNR difference between OPM_cerebellum and SQUID_cerebellum. (D) Relationship between head circumference and sensor distance ratio, showing a negative relationship. (E) Relationship between sensor distance ratio and SNR difference, showing a positive relationship. In panels, (D) and (E), each point represents a cerebellar region, and the solid line indicates the linear regression fit. (F) Mediation analysis showing the a-path (head circumference -> distance ratio), b-path (distance ratio -> SNR difference), and direct effect. (G) Indirect path (a × b) across regions. (H) Explained variance (R²) of the mediation model across cerebellar regions. Black bins indicate regions that did not reach statistical significance after FDR correction.

The dashed horizontal line indicate the median SNR of the parietal cortex in SQUID_clinical, used as a reference for best-case superficial cortical detectability.

Asterisks indicate statistical significance after FDR correction (* < 0.05, ** < 0.01, ***< 0.001). The boxplots show the SNR for each region, in the following order: lobules I-III, IV, V, VI, Crus I, Crus II, VIIb, VIIIa, VIIIb, IX, X, vermis VI, VII, VIII, IX, and X.

3.3. Distance effect on SNR for OPM_cerbellum and SQUID_cerebellum

Across cerebellar regions, head circumference showed a negative correlation with sensor distance ratio (Fig. 5D; r = −0.56, p < 0.001), indicating that larger head size reduced the relative proximity advantage of OPM sensors. The sensor distance ratio was positively correlated with SNR difference (Fig. 5E; r = 0.53, p < 0.001), consistent with greater SNR benefit when OPM sensors were relatively closer to the cerebellum.

Mediation analysis (Fig. 5F) revealed significant indirect effects (a × b) in most cerebellar regions after FDR correction, with the exception of the anterior lobes (I-III, IV, V) and Crus I, (Fig. 5G). These findings indicate that the head size influences cerebellar SNR difference primarily through its effect on relative sensor distance, particularly in posterior regions where OPM advantage was largest.

Direct effects of head circumference on SNR difference (c’ path), after controlling for distance, were generally small or non-significant, consistent with distance acting as the primary mediator of head size effects on signal detectability.

Some regions such as I-III, IV, V, VI, and Crus I did not show statistically significant indirect effects after FDR correction, despite relatively high R2 values in the regression models (Fig. 5H). Notably, the b-path was negative in a few regions (e.g., IV and V), which deviates from the expected direction. While these may reflect local anatomical or modeling differences, their non-significance after correction suggests that we cannot draw firm conclusions from these specific deviations.

Overall, these results support a mediated relationship where individuals with smaller heads (greater distance asymmetry between OPM and SQUID) tend to exhibit stronger SNR benefit from OPMs in posterior cerebellar regions.

3.4. Orientation effect on SNR difference for OPM_EEG-EEG_clinical

Folded angle maps for the cerebrum and cerebellum are shown in Fig. 6A and B, together with scatter plots illustrating the relationship between folded angle and SNR difference. Substantial intersubject variability in folded angle was observed, particularly in the medial temporal lobe in the cerebrum (Fig. 6A), as well as across most cerebellar regions (Fig. 6B).

Fig. 6.

Orientation effects on SNR difference in the OPM_EEG configuration.

Folded angle maps for the cerebrum (A) and cerebellum (B) in the OPM_EEG layout. Right panels show scatter plots illustrating the relationship between folded angle and SNR difference across regions. Solid lines indicate linear regression fits.

Linear regression analysis demonstrated a stronger relationship between folded angle and SNR difference in the cerebrum (slope = 0.64, p < 0.001) than in the cerebellum (slope = 0.23, p < 0.001). These results indicate that source orientation plays a larger role in determining detectability differences in the cerebrum, whereas in the cerebellum, depth and the combined effects of folding-related geometric and orientation dispersion dominate OPM sensitivity when OPM sensors are placed at clinical EEG electrode positions.

ROI-wise surface area together with regional summary statistics for SNR and folded angle are provided in Supplementary Table 2.

4. Discussion

In this study, we combined individualized cerebellar modeling using ARCUS with modality-specific forward modeling to evaluate signal detectability in the cerebellum across clinical SQUID-MEG, EEG and simulated OPM configurations in a cohort of patients with epilepsy. By quantifying SNR and SNR difference across modalities and sensor layouts, we provide a population-level assessment of how cerebellar anatomy, source geometry, and sensor configuration jointly shape the sensitivity of various non-invasive electrophysiological measurements across different modalities. Our results demonstrate that cerebellar detectability is strongly constrained in routine clinical settings by the neuroanatomy but can be selectively enhanced through cerebellar-optimized OPM configurations, particularly in posterior lobes.

4.1. Cerebellar functional organization and relevance to epilepsy

The cerebellum exhibits a well-established functional organization along the anterior-posterior axis. In the healthy human brain, anterior lobes are predominantly associated with sensorimotor processing (Grodd et al., 2001; Schmahmann and Sherman, 1998), whereas posterior lobules are more strongly linked to high-order cognitive and affective functions (Buckner et al., 2011; King et al., 2019; Schmahmann and Pandya, 1997; Stoodley and Schmahmann, 2009). This large-scale functional differentiation provides a useful anatomical framework for interpreting region-specific cerebellar signals in electrophysiological studies, independent of disease context.

In epilepsy, most evidence for cerebellar involvement in human brain function has come from structural (Allen et al., 2019; Lawson et al., 2000; Sandok et al., 2000), metabolic (Bohnen et al., 1998; Laich et al., 1997; Seto et al., 1997; Shin et al., 2002; Sone et al., 2022), and postmortem investigations (Crooks et al., 2000). Despite this converging evidence, direct non-invasive electrophysiological characterization of cerebellar activity has remained limited (Okada et al., 2020), possibly due to technical constrains on signal detectability. This gap motivates the present focus on cerebellar SNR characteristics across modalities and sensor configurations.

4.2. ARCUS as a framework for cerebellar electrophysiology

There are many challenges in non-invasive cerebellar electrophysiology (Samuelsson et al., 2020a, 2020b; Samuelsson et al., 2026). The primary challenge is that cerebellar foliation occurs at spatial scales comparable to, or smaller than, typical clinical MRI resolution, resulting in pronounced orientation dispersion and signal cancellation at the spatial distance relevant for MEG and EEG measurements. Under these conditions, a folium-resolved surface representation cannot be reconstructed due to the current resolution in in-vivo MRI, which is a prerequisite for MEG and EEG source modeling. ARCUS was therefore developed to provide a geometrically informed reconstruction of the cerebellar cortex suitable for electrophysiological source modeling. Importantly, under such extreme folding and orientation dispersion, vertex-level sensitivity estimates do not correspond to anatomically or physiologically independent units in the cerebellum. Individual vertices represent discretization points of a highly folded and spatially distributed current system, for which local SNR values are strongly influenced by cancellation effects and local geometry. As a result, sensitivity estimates gain interpretability only when summarized within anatomically defined cerebellar regions, where net signal contributions can be related to lobular organization and clinical relevance.

A key feature of ARCUS is the use of a high-resolution ex-vivo cerebellar template to preserve the native foliation geometry that is otherwise unresolved in standard MRI. By diffeomorphically registering this template to individual anatomy, ARCUS provides a representation of the densely folded cerebellar cortex that preserves folding geometry at a scale relevant for electromagnetic forward modeling. Importantly, the ex-vivo template is not used to impose subject-specific foliar detail beyond what is supported by the individual lobular segmentation data, but rather to inform a source-space representation that respects the intrinsic folding geometry of the cerebellar cortex. In this framework, vertices in the ARCUS-derived source-space should not be interpreted as representing individual folia or local cortical patches, but instead as discretization points of a distributed cortical geometry whose fine-scale folding is preserved in aggregate. This representation is well aligned with the biophysical reality that cerebellar electromagnetic signals arise from spatially distributed current sources whose net contribution reflects both geometry and orientation dispersion.

This approach fundamentally distinguishes ARCUS from atlas-based cerebellar parcellation methods such as SUIT or FreeSurfer-based cerebellar parcellations (Buckner et al., 2011; Diedrichsen, 2006; Diedrichsen et al., 2009, 2011; Faber et al., 2022; Han et al., 2020; Lyu et al., 2024). Whereas those methods provide anatomically defined, and more recently functionally defined (Nettekoven et al., 2024), lobule-level parcellations that are highly valuable for morphometric and volumetric analyses, they do not provide a cerebellar cortical reconstruction that adequately captures the cortical folding which can be used to model neuronal currents relevant for electromagnetic source modeling. In contrast, ARCUS provides an explicit cortical geometry derived from a high-resolution cerebellar template, enabling a source discretization that is directly informed by cerebellar folding and source organization, and is therefore naturally compatible with forward models used in MEG and EEG.

The theoretical feasibility of modeling cerebellar activity with MEG and EEG was first established in our earlier work (Samuelsson et al., 2020b), which demonstrated that cerebellar activity can generate measurable electromagnetic signals despite substantial foliation-induced cancellation, using a high-resolution ex-vivo cerebellar template. Building on this theoretical foundation, subsequent work showed that such high-resolution cerebellar surfaces can be adapted for electrophysiological source modeling in controlled research settings (Alho et al., 2023).

The present study extends these prior efforts by demonstrating that the ARCUS framework enables cerebellar source modeling robustly and at scale in heterogeneous real-world clinical MRI datasets. By systematically applying ARCUS to routine clinical MRI data acquired under variable imaging conditions, we show that anatomically valid cerebellar segmentations and corresponding source spaces can be obtained in the vast majority of patients (95%), despite variability in field strength and slice thickness. This robustness is essential for translating cerebellar electrophysiology from proof-of-principle studies to large-scale clinical and population-based investigations. Although we did not quantitatively assess folium-level reconstruction accuracy, visual inspection against individual MRI confirmed anatomically plausible lobular segmentation in the vast majority of cases, supporting the suitability of the derived source spaces for forward modeling relevant for the spatial scale considered here.

4.3. Cerebellar SNR across modalities and sensor configurations

Our SNR analyses suggest that, in routine clinical SQUID-MEG and EEG recordings, detection of cerebellar activity is challenging (Fig. 4A). When OPM sensors were simulated at scalp locations corresponding to the SQUID array—thereby reducing source-to-sensor distance while preserving geometry and coverage—no meaningful improvement in cerebellar SNR was observed (Fig. 4F). In contrast, cerebellar-optimized OPM configurations yielded substantial SNR improvements in posterior cerebellar regions (Fig. 5C). Together, these results demonstrate that targeted coverage and geometry, rather than sensor proximity alone, are critical for improving detectability of deep and highly folded structures such as the cerebellum.

Importantly, this pattern was not unique to the cerebellum. Several deep cerebral regions—including the medial temporal lobe, cingulate cortex, and insula also failed to show meaningful SNR improvement when OPM sensors were placed closer to the scalp using SQUID-matched geometry (Fig. 3F). This convergence across cerebellar and cerebral regions provides and important insight: for deep sources, scalp-source distance is a dominant limiting factor, and reducing sensor distance by bringing sensors closer to the scalp alone does not significantly improve SNR (Boto et al., 2016; Iivanainen et al., 2017).

One plausible explanation is that moving sensors closer increases both signal and background brain-noise contributions in a comparable manner for deep sources, producing minimal net SNR gain (Brookes et al., 2022). In the present framework, SNR is determined by the forward model together with the assumed noise structure. Changes in SNR across sensor layouts therefore do not reflect a preferential amplification of neural signal relative to noise, but rather arise from differences in how source geometry, orientation, and spatial dispersion are mapped onto sensors through the forward solution. Under these conditions, simply shifting a fixed geometry closer to the scalp does not overcome the fundamental depth-related limitations.

By contrast, the cerebellar-optimized OPM configurations explicitly targets the region of interest by altering coverage, while modeling the effect of geometry via forward modeling, thereby improving alignment between sensor sensitivity profiles and cerebellar source distributions. This within-modality gain is summarized in Fig. 7A, which highlights posterior cerebellar regions where a cerebellar-optimized OPM layout yields higher SNR than a geometry-matched SQUID configuration. The resulting SNR gains—most evident in posterior lobules—highlight that the principal advantage of OPMs lies not in proximity alone, but in the ability to design sensor layouts optimized for specific deep structures, which is not possible in fixed-helmet SQUID systems.

Fig. 7.

Complementary summary views of cerebellar SNR.

(A) summarizes the effect of sensor geometry optimization within MEG in the comparison between OPM_cerebellum and SQUID_cerebellum, whereas (B) provides a qualitative visualization of complementary sensitivity across MEG and EEG modalities.

4.4. Orientation effects in the OPM-EEG comparison

Given the recognition that depth imposes a dominant constraint on detectability, we examined the contribution of source orientation to SNR differences when OPM and EEG were compared under identical sensor coverage (Fig. 2B and D). In this configuration, folded angle emerged as a statistically significant determinant of SNR difference, with different magnitudes in the cerebrum and cerebellum.

The relationship between folded angle and SNR difference was stronger in the cerebrum (Fig. 6A) than in the cerebellum (Fig. 6B). Notably, the medial temporal lobe exhibited the greatest intersubject variability in folded angle. This variability likely reflects substantial interindividual differences in local cortical geometry, but it may also be amplified by the relatively coarse ROI definition used here; collapsing multiple medial temporal substructures into a single ROI may pool heterogeneous folding patterns and inflate apparent orientation variability.

In the cerebellum, folded angle variability was also substantial across regions, yet its contribution to SNR difference was more modest. This attenuation is consistent with greater depth and orientation dispersion of cerebellar sources, suggesting that geometric and depth-related factors dominate detectability constraints and limit the extent to which orientation alone can drive regional SNR differences. A further point to consider is that the folded angle analysis in Fig. 6B reflects the ARCUS-derived macroscale cerebellar surface representation obtained from routine MRI, rather than the full microscopic distributions of local folial orientations. ARCUS provides an anatomically informed, template-fitted cortical sheet that preserves lobule-scale folding geometry, but does not reconstruct subject-specific folial contours point-by-point from clinical MRI. Accordingly, the folded angle values shown here should not be interpreted as excluding the presence of more near-radial local orientations within the cerebellar folia; rather, such local extremes are not expected to be fully represented in this macroscale source model.

Importantly, while several posterior regions showed higher SNR with OPM than EEG (Fig. 4G), EEG retained higher sensitivity in other cerebellar regions. This complementary pattern indicates that OPM and EEG provide distinct but overlapping sensitivity profiles (Seedat et al., 2024), supporting their combined use to extend non-invasive evaluation beyond the cerebral cortex. To emphasize this multimodal complementarity, Fig. 7B provides a qualitative summary indicating, at each cerebellar vertex, the modality (SQUID_cerebellum, OPM_cerebellum, or EEG-clinical) within the highest estimated sensitivity. Notably, no single modality dominates across the cerebellum, underscoring the value of integrating EEG and MEG to achieve broader electrophysiological coverage of cerebellar regions. This categorical representation is intended as a conceptual visualization rather than a basis for quantitative or statistical comparison and served to illustrate how multimodal integration can mitigate the inherent limitations of individual modalities in probing deep and highly folded brain structures.

4.5. Anatomical determinants mediating OPM-SQUID SNR differences

We also examined whether anatomical factors mediate the relative SNR advantage of OPM-based configurations over SQUID systems in the cerebellum. By explicitly modeling sensor distance and head size, we demonstrate that anatomical factors play a key mediating role in cerebellar SNR differences between OPM and SQUID configurations.

Mediation analysis revealed that the effect of head size on SNR difference was significantly mediated by relative sensor distance in most posterior cerebellar regions (Fig. 5F–H), whereas the anterior lobes (I–III, IV, V) and Crus I did not show significant mediation effects. These findings indicate that individuals with smaller heads—who exhibit greater asymmetry in sensor proximity between OPM and SQUID systems—derive a larger SNR benefit from cerebellar-optimized OPM layouts. This reinforces the conclusion that depth-related limitations can be partially mitigated when sensor geometry and coverage are adapted to individual anatomy, particularly for posterior cerebellar regions where OPM advantage is most evident.

These observations have important implications for clinical translation. Conventional SQUID-MEG relies on fixed helmet geometries that are optimized primarily for adult head sizes, which can further constrain sensitivity for deep sources in smaller heads. BabyMEG systems provide an important proof of concept that MEG helmets optimized for pediatric head geometries can improve MEG measurements and signal quality (Okada et al., 2016), but such systems remain specialized and are not widely available. In contrast, OPM-based systems offer a flexible alternative, enabling subject-specific sensor placement that can accommodate a broad range of head sizes while targeting regions such as the posterior fossa.

Together, these findings suggest that anatomically adaptive sensor layouts are critical for improving cerebellar detectability, and that OPM-MEG may be particularly advantageous in populations for whom fixed-helmet system are most constrained.

4.6. Limitations and future directions

Several limitations should be acknowledged. First, OPM results in this study were derived from simulations rather than patient recordings, and noise estimates were extrapolated from SQUID and EEG data. Although this approach was supported by a validation experiment in a healthy individual (Supplementary material, and Supplementary Fig. 1), the acquisitions were not fully equivalent: OPM and SQUID recordings were obtained in different magnetic shield room environments (single-layer versus three-layer), employed different postprocessing strategies, and involved different sensor counts. These differences limit direct quantitative generalization and motivate future work incorporating empirically measured OPM noise characteristics in clinical populations.

Second, the SNR framework relies on forward modeling, which necessarily simplifies the relationship between neural signals and noise. While this enables systematic comparison across modalities and sensor configurations, it does not capture all physiological and environmental noise contributions that may vary with tasks and recording conditions. Future studies combining subject-specific forward models with task-dependent OPM recordings will be important to refine estimates of cerebellar detectability under clinical realistic conditions.

Third, the cerebellar-optimized OPM configuration examined here was designed to enhance posterior fossa coverage but was not explicitly optimized using detailed cerebellar anatomy, functional organization, or expected source extent. More principled optimization strategies incorporating cerebellar lobular anatomy, orientation distributions, and anticipated spatial extent may yield additional improvements and represent an important direction for future cerebellar-focused OPM system design.

Finally, although we visually validated the anatomical plausibility of ARCUS-derived cerebellar segmentation against individual MRI at the lobular level, we did not directly assess geometric accuracy at the level of individual folia. Such validation would require ultra-high-resolution ground truth data not available in routine clinical imaging. Importantly, the present study was designed to evaluate the suitability of the ARCUS-derived source space for forward modeling and SNR estimation at the spatial scale relevant to MEG and EEG, rather than to validate folium-level surface reconstruction accuracy.

5. Conclusions

Cerebellar optimized OPM configurations increased SNR particularly in posterior cerebellar regions, whereas proximity alone did not yield meaningful improvements for deep sources. Integration of individualized cerebellar modeling using ARCUS with scalp-mounted OPM arrays may help extend non-invasive electrophysiological coverage beyond the cerebral cortex. Together, these advances provide a realistic path toward improving the assessment of cerebellar involvement in epilepsy, a region that has traditionally been difficult to evaluate in clinical practice.

Supplementary Material

Supplementary material associated with this article can be found, in the online version, at doi:10.1016/j.neuroimage.2026.121930.

Abbreviations:

ARCUS

Automatic Reconstruction of Cerebellar Cortex from Standard MRI Using Diffeomorphic Registration of a High-Resolution Template

BEM

Boundary element model

EEG

Electroencephalography

FDR

False discovery rate

MEG

Magnetoencephalography

MRI

Magnetic resonance imaging

OPM

Optically pumped magnetoencephalography

ROI

Region of interest

SD

Standard deviation

SNR

Signal-to-noise ratio

SQUID

Superconducting quantum interference device

SUIT

Spatially unbiased infratentorial template

Data availability

The data are available from the corresponding author upon reasonable request and with appropriate institutional approvals.

Code availability

Analysis scripts used for anatomically informed forward modeling and signal-to-noise ratio estimation are available from the corresponding author upon reasonable request.