Seizure‐onset zone lateralization in temporal lobe epilepsy using 7T rs‐fMRI: Direct comparison with 3T rs‐fMRI

Abstract

Objective

Resting‐state functional magnetic resonance imaging (rs‐fMRI) at ultra‐high field strengths (≥7T) is known to provide superior signal‐to‐noise to comparable acquisitions at lower field strengths. In this study, we provide a direct comparison of the seizure onset‐zone (SOZ) lateralizing ability of 7T rs‐fMRI and 3T rs‐fMRI.

Methods

We investigated a cohort of 70 patients with temporal lobe epilepsy (TLE). A paired cohort of 19 patients had 3T and 7T rs‐fMRI acquisitions for direct comparison between the two field strengths. Forty‐three patients had only 3T, and eight patients had only 7T rs‐fMRI acquisitions. We quantified the functional connectivity between the hippocampus and other nodes within the default mode network (DMN) using seed‐to‐voxel connectivity, and measured how hippocampal–DMN connectivity could inform SOZ lateralization at 7T and 3T field strengths.

Results

Differences in hippocampal–DMN connectivity ipsilateral and contralateral to the SOZ were significantly higher at 7T (Cohen's d = 0.51, p = 0.008) than at 3T (Cohen's d = 0.26, p = 0.68) when measured in the same subjects. We found that SOZ lateralization was superior at 7T (receiver‐operating characteristic area under the curve [ROC AUC] = 0.97, 95% confidence interval [CI]: 0.92–1.00) than 3T (ROC AUC = 0.67, 95% CI: 0.36–0.98), for the same subjects scanned at both field strengths. Our findings were reproduced in extended cohorts of subjects scanned at either 3T or 7T. Our rs‐fMRI findings at 7T, but not 3T, are consistent (Spearman ρ = 0.65, p = .01) with clinical fluorodeoxyglucose positron emission tomography (FDG‐PET) lateralizing hypometabolism.

Significance

We show superior SOZ lateralization in patients with TLE when using 7T relative to 3T rs‐fMRI, supporting the adoption of high field strength functional imaging in the epilepsy presurgical evaluation.

Key points.

• 7T resting‐state functional magnetic resonance imaging (rs‐fMRI) provides superior seizure‐onset zone (SOZ) lateralization compared to 3T in patients with temporal lobe epilepsy (TLE).

• Hippocampal–DMN (default mode network) connectivity differences are significantly higher at 7T than at 3T.

• 7T rs‐fMRI lateralization findings align with fluorodeoxyglucose positron emission tomography (FDG‐PET) hypometabolism, but 3T does not.

• Results were reproduced in independent patient cohorts, reinforcing the reliability of 7T over 3T.

1. INTRODUCTION

Functional magnetic resonance imaging (fMRI) is the main non‐invasive approach for assessing whole brain function in vivo. In epilepsy, task‐based fMRI at field strengths of 1.5‐Tesla (T) and 3T are used clinically for language lateralization prior to surgical resection. ¹ Resting‐state fMRI (rs‐fMRI), however, remains yet to be clinically translated to presurgical evaluation of epilepsy, despite studies at 3T demonstrating potential for seizure‐onset zone (SOZ) lateralization. ² , ³ , ⁴ One well‐recognized limitation that impedes translation is the relatively moderate test–retest reliability of rs‐fMRI, arising from factors such as head motion, physiological fluctuations, and differences in cognitive or vigilance states. Small sample sizes and low statistical power also remain limiting factors, making it challenging to detect robust, clinically relevant findings at a single subject level. ⁵

Despite these concerns, many rs‐fMRI studies in epilepsy at 3T have detected functional changes that suggest network reorganization. ² , ³ , ⁴ , ⁶ , ⁷ For example, decreased ipsilateral hippocampal connectivity to the default mode network (DMN) remains a commonly reproduced finding in case–control studies of temporal lobe epilepsy (TLE). ⁸ , ⁹ , ¹⁰ , ¹¹ , ¹² , ¹³ However, the relatively low signal‐to‐noise ratio (SNR) and contrast‐to‐noise ratio (CNR) at lower field strengths can hinder the detection of subtle connectivity differences and may contribute to lower reproducibility. Moving to higher field strengths such as 7T can help address this limitation: the SNR and CNR of the blood oxygen level–dependent (BOLD) signal increase with higher field strength, a well‐established finding supported by prior work. ¹⁴ This enhancement improves sensitivity to small BOLD fluctuations and potentially mitigates some sources of variability (e.g., partial volume effects in smaller structures like the hippocampus), although subject motion and vigilance‐related changes remain factors regardless of field strength.

At 7T, rs‐fMRI studies have also uncovered functional network reorganization in epilepsy, ¹⁵ , ¹⁶ , ¹⁷ but none have directly compared the utility of 3T vs 7T fMRI in describing the underlying epileptogenic networks, or lateralizing the SOZ. This is in contrast to the many studies comparing the clinical utility of 7T and 3T structural neuroimaging in epilepsy. ¹⁸ , ¹⁹ , ²⁰ , ²¹ , ²² , ²³ Thus, whether 7T rs‐fMRI might offer superior reproducibility or improved accuracy in lateralizing the SOZ when directly compared to 3T rs‐fMRI remains an open question.

In this study, we aimed to address the gap in knowledge regarding the clinical utility of 7T rs‐fMRI when compared to 3T rs‐fMRI in TLE. To do so, we asked whether 7T rs‐fMRI could provide better lateralization than 3T rs‐fMRI of the SOZ in focal TLE. To answer this question, we investigated (1) a cohort of TLE patients with paired 7T and 3T rs‐fMRI acquisitions, allowing us to directly compare findings across field strengths, and (2) an extended cohort of patients with only 3T or 7T rs‐fMRI, allowing for further validation of the findings from the paired cohort. For each patient, and at both field strengths, we measured the hippocampal functional connectivity to other nodes within the DMN. We then compared how the hippocampal–DMN connectivity ipsilateral to the SOZ differed from the connectivity contralateral to the SOZ, a common approach to study SOZ laterality in the TLE literature. ⁸ , ⁹ , ¹⁰ , ¹¹ , ¹² , ¹³ Finally, we compared the utility of 3T and 7T rs‐fMRI in lateralizing the SOZ at a single subject level, and compared the performance of each field strength with the lateralization of the SOZ by quantitative FDG‐PET. Our findings provide evidence for superior SOZ lateralizing ability by 7T rs‐fMRI, supporting adoption of high field strength functional neuroimaging in the epilepsy presurgical evaluation.

2. METHODS

2.1. Subject inclusion

Data acquisition for this study was approved by the institutional review board of the University of Pennsylvania. Written informed patient consent was obtained from all participants. Patients with drug‐resistant unilateral TLE who underwent neuroimaging during presurgical evaluation between the years 2016 and 2022 at our Level 4 epilepsy center were included. Neuroimaging included structural and resting state functional magnetic resonance imaging (fMRI). Subject inclusion criteria consisted of failure of at least two or more anti‐seizure medications, and seizures originating from the temporal lobe and associated structures (e.g., hippocampus and amygdala). The Penn Epilepsy Surgical Conference (PESC) determined the location of seizure focus for all patients using a multidisciplinary approach, including clinical, neuroimaging, and neurophysiological factors such as seizure semiology, MRI, FDG‐PET, magnetoencephalography (MEG), scalp electroencephalography (EEG), and intracranial EEG (iEEG). A cohort of 70 patients was used in this study. Nineteen patients received both 3T and 7T neuroimaging as part of a research‐specific imaging protocol. Thirty‐nine received only 3T and eight received only 7T neuroimaging, also as part of research‐specific protocols. Demographic information, MRI lesional status, presence of mesial temporal sclerosis (MTS), final lateralization, and 12‐month Engel surgical seizure outcomes, ²⁴ where available, are recorded in Table 1. For non‐MTS lesional subjects, lesions seen on MRI included: encephalocele, encephalomalcia, polymicrogyria, heterotopia, past stroke, focal cortical dysplasia (FCD), and neoplasms.

TABLE 1.

Demographic and clinical information for paired 3T–7T cohort as well as extended 3T and 7T cohorts.

Subject demographics
Characteristic	Group
	Paired 3T–7T	Extended 3T	Extended 7T
Number of subjects	19	62	27
Age	35 ± 11	38 ± 13	36 ± 11
Female	12	34	15
MRI lesional	12	40	17
MTS	8	22	11
Lateralization
Left	12/19	37/62	17/27
Right	7/19	25/62	10/27
Surgical intervention	9	25	11
Engel I	6/9	16/25	8/11
Engel II–IV	3/9	9/25	3/11
Imaging protocol version
A (3T)	5/19	13/62	—
B (3T)	14/19	49/62	—
C (7T)	4/19	—	10/27
D (7T)	15/19	—	17/27

Abbreviations: Engel, Engel surgical outcome; MTS, mesial temporal sclerosis.

2.2. Image acquisition

Resting‐state fMRI images at 3T and 7T were acquired using four protocols. At 3T, Protocol A utilized a Siemens MAGNETOM TrioTim, capturing images with a 3.0 mm isotropic voxel size, echo/repetition time (TE/TR) of 30/500 ms, a multiband factor of 6, and a 7‐min acquisition time. Protocol B, with a Siemens MAGNETOM PrismaFit, had a 2.0 mm voxel size, TE/TR of 37/800 ms, a multiband factor of 6, and a 6‐min acquisition. Both involved resting‐state scans with eyes closed and T1 weighted images via a 208‐slice magnetization‐prepared rapid gradient echo (MPRAGE) sequence, TE/TR of 2.24/2400 ms, inversion time (TI) of 1060 ms, field of view (FOV) of 256 mm, and a 0.8 mm voxel size. At 7T, a Siemens MAGNETOM Terra scanner was used. 7T Protocol C involved 2.0 mm voxel size, TE/TR of 23.6/1000 ms, and a 7‐min session; Protocol D featured a 1.5 mm voxel size, TE/TR of 20/1000 ms, both with a multiband factor of 6, and a 10‐min session. T1 images for C were captured with a 176 slice MPRAGE sequence, TE/TR of 4.4/2800 ms, TI of 1500 ms, and a 0.8 mm voxel size; and for D, a 160 slice magnetization‐prepared 2 rapid gradient echo (MP2RAGE) sequence, TE/TR of 1.86/5000 ms, TI1/TI2 of 700/2700 ms, and a 1 mm voxel size.

2.3. Functional imaging processing

We used fMRIPrep ²⁵ to perform brain extraction and segmentation of individual T1‐weighted (T1w) images, registration of task fMRI BOLD volumes to individual T1w and Montreal Neurological Institute (MNI) template space, and time‐series confound estimation. We used the fMRIPrep output data as our input to the xcpEngine post‐processing pipeline for confound regression, demeaning, detrending, and temporal filtering. ²⁶ The same pipeline was used for 3T and 7T datasets regardless of protocol. The only exception is that for the 7T MP2RAGE structural images from Protocol D, we used the multiplication of the INV2 and UNI image as the T1w input to fMRIprep, allowing for improved brain extraction. We found no differences in motion between subjects scanned at 3T or 7T (Figure S1). See Methods S1 for specific details on the pre‐processing pipeline used in this study.

2.4. tSNR analysis

To determine if differences in temporal signal to noise ratio (tSNR) could be influencing our results at 3T and 7T, we measured tSNR in the fMRI acquisitions. Because of the differences in protocols, we used a corrected tSNR calculation based on the voxel size, repetition time, and number of timepoints, as reported previously, ²⁷ , ²⁸ resulting in the following equation:

$$\text{tSNR}=\frac{{\mu }_{\text{BOLD}}}{{\sigma }_{\text{BOLD}}}\frac{\sqrt{N}}{V\sqrt{\mathrm{TR}}}$$

Where ${\mu }_{\text{BOLD}}$ is the mean BOLD signal, ${\sigma }_{\text{BOLD}}$ is the standard deviation (SD) of the BOLD signal, $N$ is the number of timepoints, $V$ is the voxel volume, and TR is the repetition time.

We measured the tSNR on the minimally preprocessed fMRI images, consisting of the fMRI acquisition transformed into MNI space but prior to regression of biological and motion parameters (e.g., the preprocessed BOLD in MNI space from fMRIprep prior to input to xcpEngine; see Supporting Information Methods S1 for details). Chosen regions were based on the Harvard‐Oxford cortical and subcortical atlases, ²⁹ , ³⁰ with the posterior cingulate cortex (PCC) and the medial pre‐frontal cortex (MPFC) defined as specified in: Comparing Left and Right Seed‐to‐Voxel Maps—Group Level.

2.5. Hippocampal seed‐to‐voxel connectivity

A general overview of the methods is shown in Figure 1. To measure connectivity of the hippocampus to the DMN, we used seed‐to‐voxel connectivity (Figure 1B). For each patient, we defined a mask in MNI space for the left and right hippocampi based on the Harvard‐Oxford subcortical atlas. We computed the average processed BOLD timeseries within the left and right hippocampal regions of interest (ROIs), which we defined as the left and right seeds, respectively. Then, we computed the Pearson's correlation of the timeseries of each seed ROI with the timeseries of every other voxel in the brain.

FIGURE 1.

Methodological overview. (A) Shows the unprocessed BOLD, T1‐weighted and processed BOLD images for the same participant at 3T and 7T. Functional imaging processing was carried out for both 3T and 7T with fMRIprep and XCPEngine. For the processed 3T and 7T BOLD images, the hippocampal seed‐to‐voxel connectivity was computed for the left and right hippocampi as shown in (B). The left and right seed‐to‐voxel maps were then used in subsequent analyses throughout the manuscript. BOLD, blood‐oxygen‐level dependent signal.

All connectivity analyses were performed on MNI‐coregistered images that were standardized to a 2 mm isotropic resolution as part of the functional preprocessing pipeline. This means that rs‐fMRI images from protocols with voxel sizes larger or smaller than 2 mm were resampled to 2 mm isotropic before seed‐to‐voxel calculations. Therefore all resulting seed‐to‐voxel maps are presented and analyzed in 2 mm isotropic space. The mean seed‐to‐voxel connectivity maps across the paired cohort are shown in Figure S2.

2.6. Comparing left and right seed‐to‐voxel maps—Group level

To measure whether the difference between left and right seed‐to‐voxel maps differed between left (L‐TLE) and right TLE (R‐TLE), and whether the measured difference was larger at 7T than at 3T, we used voxelwise Cohen's d maps. We calculated a Cohen's d map by comparing right seed‐to‐voxel maps with left seed‐to‐voxel maps. Negative Cohen's d values indicate that the right hippocampus is less connected to a particular voxel than the left hippocampus, whereas positive values indicate the opposite. We smoothed the Cohen's d maps with a Gaussian filter with σ = 0.5 voxels for distribution analyses, and with a filter of σ = 2 voxels for visualization and correlation analyses.

We measured the distribution of Cohen's d values in two nodes of the DMN: the PCC and the MPFC (Figure 2A). These nodes were chosen due to their prominence within the DMN, ³¹ as well as their previously reported association with TLE. ¹² , ¹³ We created PCC and MPFC masks by combining Harvard‐Oxford regions and selecting DMN ROI voxels using the Schaefer 200 atlas. ³² The PCC mask included the cingulate gyrus (posterior division) and precuneus cortex, whereas the MPFC mask included the frontal pole and superior frontal gyrus. We generated separate masks for the left and right hemispheres and measured their distributions in the combined mask.

FIGURE 2.

3T vs 7T tSNR Comparison. (A) The voxelwise mean difference in tSNR is shown between 3T and 7T acquisitions for the paired cohort. Negative (blue) values indicate a higher tSNR for 7T than for 3T and positive values (red) the opposite. (B) Shows the structural slices corresponding to those shown in (A). (C) Shows the Cohen's d between the 3T and 7T estimated tSNR across Harvard‐Oxford cortical ROIs. (D) Shows (top left) the mean cortical and subcortical tSNR values, (top right) the mean tSNR values for the PCC and MPFC in the DMN, and (bottom) the mean tSNR values across all subcortical ROIs for the paired 3T and 7T subjects. *0.01 < p ≤ .05, **0.001 < p ≤ .01, ***0.0001 < p ≤ .001. DMN, default mode network; MPFC, medial prefrontal cortex; PCC, posterior cingulate cortex; ROI, region of interest; TLE, temporal lobe epilepsy; tSNR, temporal signal‐to‐noise ratio.

2.7. Comparing left and right seed‐to‐voxel maps—Subject level

To measure the ability of 3T and 7T rs‐fMRI to lateralize the SOZ, we computed for each subject the mean value of the difference between smoothed (Gaussian filter with σ = 0.5 voxels) right and left seed‐to‐voxel hippocampal maps in the PCC. For this analysis, we restricted the measurement to the PCC because it was the region where we observed the strongest effect in the group level analysis.

2.8. Quantitative FDG‐PET analysis

FDG‐PET remains the neuroimaging clinical standard for SOZ lateralization in TLE ³³ ; therefore, we compared quantitative FDG‐PET hypometabolism to our rs‐fMRI findings at 3T and 7T. Eighteen of the 19 paired subjects had clinical FDG‐PET scans available for this analysis. For each subject, FDG‐PET scans were registered through a rigid transformation to the T1w MRI. Then we generated subject‐specific parcellations of the Desikan‐Killiany‐Tourville (DKT) atlas using FreeSurfer on the T1w image. ³⁴ From these parcellations, we measured the asymmetry index (right–left/right+left) of the FDG‐PET contrast within a temporal (DKT labels corresponding to the superior, middle and inferior temporal giri) and a mesial temporal ROI (DKT labels corresponding to the hippocampus, amygdala, parahippocampus and entorhinal cortex). The caudal aspect of the middle frontal gyrus (near the supplementary motor area) was used as a control region.

2.9. Statistical analysis

To compare the distribution of Cohen's d values in the PCC and MPFC between the L‐ and R‐TLE, we performed a permutation test of the difference between the mean of the L‐TLE and R‐TLE Cohen's d distributions. We carried out 1000 permutations of the difference between the mean Cohen's d in L‐TLE and R‐TLE, and then calculated the p‐value for the difference. ³⁵ Additional comparisons between L‐TLE and R‐TLE were done using two‐sample, two‐tailed t‐tests. tSNR analyses were done using a Mann–Whitney U test. For all analyses, statistical significance was determined at p < .05, and Bonferroni correction was used to correct for multiple comparisons where specified (p _FWER). Reported Cohen's d values are absolute values. For confidence interval (CI) estimates of the area under the receiver‐operating characteristic curve (ROC AUC), the DeLong approximation was used. ³⁶

3. RESULTS

3.1. tSNR of 7T is higher in midline and subcortical structures

We measured the temporal signal‐to‐noise ratio (tSNR) maps from the rs‐fMRI acquired at 3T and 7T in our paired cohort. Our comparisons show that, on average, the tSNR for 7T was higher for midline and deeper subcortical structures, whereas the tSNR for 3T was higher for lateral cortical structures (Figure 2A). This was the case also when the Cohen's d between the tSNR at 3T and 7T was calculated for Harvard‐Oxford cortical and subcortical ROIs, showing that midline and ROIs had higher Cohen's d for 7T (Figure 2C). Statistical comparison of the tSNR in subcortical ROIs showed that it was significantly higher for 7T than for 3T (Cohen's d = 1.30, p _FWER < .001) (Figure 2D). There were no statistically significant differences between 3T and 7T for cortical tSNR (Cohen's d = 0.130, p _FWER = .860). Within specific subcortical ROIs (Figure 2D), there was a statistically significant higher tSNR for 7T only in the thalamus (Cohen's d = 1.52, p _FWER < .001), but all other regions except the nucleus accumbens had higher tSNR on average for 7T, including the amygdala (Cohen's d = 0.813, p = .030) and hippocampus (Cohen's d = 0.677, p = .055). Within cortical DMN ROIs, no statistically significant differences were seen between 3T and 7T, but on average, 7T had higher tSNR in the PCC (Cohen's d = −0.564, p = .178). Overall, our results suggest increased subcortical and midline cortical tSNR at 7T.

3.2. Lateralizing group‐level effects are stronger at 7T than 3T

To assess effect size differences between 3T and 7T rs‐fMRI in quantifying hippocampal–DMN connectivity in TLE, we measured group‐level Cohen's d maps between left and right seed‐to‐voxel maps (Figure 3). At 7T (Figure 3C,D, Figure 4A), significant functional connectivity differences were observed between the left TLE (L‐TLE) and right TLE (R‐TLE) within the PCC and MPFC, with respective Cohen's d differences of 0.51 (pFWER = .008) and 0.30 (pFWER = .042). In contrast, at 3T (Figure 3C,D, Figure 4A), differences were less pronounced and not statistically significant, with Cohen's d values of 0.26 (pFWER = .680) in the PCC and 0.35 (pFWER = .160) in the MPFC. The extended 3T cohort showed no significant differences, whereas the extended 7T cohort exhibited a difference in the PCC (Cohen's d = 0.42, pFWER = .048), although less than the original paired 7T cohort. The supplementary motor area (SMA) served as a control region, showing no significant differences between L‐TLE and R‐TLE.

FIGURE 3.

Smoothed voxelwise Cohen's d values: (A) Shows the sagittal, coronal, and axial slices that are used in all the other panels of the figure. The red and blue lines represent the locations of the right and left sagittal slices, respectively. (B) Shows the sagittal view of the posterior cingulate cortex (PCC) and the medial prefrontal cortex (MPFC) ROIs used in the study. (C, D) Show the smoothed (spatial Gaussian filter with $\sigma =2$) voxel‐wise Cohen's d maps between the grouped left and right seed‐to‐voxel maps for the left and right TLE, respectively. The first, second, third, and fourth rows of (C) and (D) show the maps corresponding to the paired 3T, the extended 3T, the paired 7T, and the extended 7T cohorts, respectively. 3T—paired 3T cohort, 7T—paired 7T cohort, 3T‐ext.—extended 3T cohort, 7T‐ext.—extended 7T cohort. TLE, temporal lobe epilepsy.

FIGURE 4.

Voxelwise Cohen's d distribution in default mode network (DMN) nodes: (A) Top row, shows the distribution of voxelwise Cohen's d values (from Figure 3) within the posterior cingulate cortex (PCC) ROI for (from left to right) the paired 3T, the extended 3T, the paired 7T, and the extended 7T cohorts. The middle and bottom rows show the same distributions but for the medial prefrontal cortex (MPFC) and the supplementary motor area (SMA), the latter used as a control region. Statistically significant differences between distributions are specified in the figure. (B, C) show the voxel‐wise Pearson correlation between pairs of Cohen's d maps for L‐TLE (B) and R‐TLE (C). Top, middle, and bottom show the correlation within the PCC, MPFC, and a whole brain mask. All correlations were statistically significant (p _FWER <.05) unless otherwise specified. NS, not statistically significant. L‐TLE, left temporal lobe epilepsy; ROI, region of interest; R‐TLE, right temporal lobe epilepsy.

3.3. Small sample sizes at 7T replicate findings in large samples at 3T

To determine if increasing the sample size at 3T makes findings more consistent with those seen at 7T, we computed voxel‐wise correlations between Cohen's d maps of the paired and extended 3T and 7T cohorts (Figure 4B,C). In R‐TLE (Figure 4C), we saw in the PCC that the extended 3T cohort had a much better spatial correlation with the paired (Pearson's r = .62, p _FWER < .001) and extended (Pearson's r = .63, p _FWER < .001) 7T cohorts, compared to the original correlation between the paired 3T and 7T cohorts (Pearson's r = −.008, p _FWER = .10). That is, as we increased the number of subjects in the R‐TLE group scanned at 3T, we saw findings more consistent with those seen at 7T. This pattern was also observed in the MPFC and whole brain, although to a lesser extent. For L‐TLE (Figure 4B), correlations between the extended 3T cohort and the paired and the extended 7T cohorts remained low if the original paired 3T–7T correlation was low (as in the PCC), or high if the original paired 3T–7T correlation was high (as in the MPFC). For the whole brain correlation, we saw a pattern identical to that of R‐TLE. The findings in the extended 3T cohort suggest that larger sample sizes at 3T lead to spatial distributions of effects that are more consistent with those seen at a much smaller sample size at 7T.

3.4. Hippocampal–DMN connectivity lateralizes the SOZ in TLE better at 7T than at 3T

To quantify the ability of 3T and 7T rs‐fMRI in lateralizing the SOZ we used the hippocampal–PCC connectivity by computing the mean value of the difference between smoothed right and left seed‐to‐voxel hippocampal maps within the PCC (Figure 5). At 7T (Figure 5A), the right–left hippocampus–PCC connectivity difference was significantly higher for L‐TLE than for R‐TLE (p < .001, Cohen's d = 2.42). This demonstrates hypoconnectivity between the left hippocampus and the PCC in L‐TLE, and hypoconnectivity between the right hippocampus and the right PCC in R‐TLE, consistent with group‐level findings. For the same subjects at 3T (Figure 5B), the difference was not statistically significant (p = .206, Cohen's d = 0.62). Using the right–left hippocampus–PCC connectivity difference as a classification metric for separating L‐TLE and R‐TLE, we achieved superior lateralization of the SOZ at 7T (ROC AUC = 0.97, 95% CI: 0.92–1.00) compared to 3T (ROC AUC = 0.67, 95% CI: 0.36–0.98). In addition, we tested whether this effect was consistent across the different 3T and 7T scanning protocols used in the study, and our results showed consistent results across all protocols (Figure S3). We also repeated this analysis using subject‐specific hippocampal segmentations, which also showed consistent findings (Methods S1; Figure S4). Finally, to test whether this effect was specific to the DMN, repeated this analysis across all the Harvard‐Oxford cortical ROIs as well as whole hemispheres, and found that regions belonging to the DMN (PCC, MPFC, and lateral temporal lobes) had the highest effect size, and that this effect was highest at 7T (Figures S5 and S6). These findings demonstrate superior SOZ lateralizing ability using the hippocampal–DMN connectivity at 7T compared to 3T.

FIGURE 5.

Comparing left and right seed‐to‐voxel maps at a subject level: (A, B) correspond to a subset of the paired 7T and 3T cohorts, respectively. On the left, the difference between the right and left (right minus left) hippocampal seed‐to‐voxel maps for the same six left TLE subjects is shown at 7T (A) and 3T (B) field strengths. In the middle, the difference between right and left (right minus left) hippocampal seed‐to‐voxel maps for the same six right TLE subjects is shown at 7T (A) and 3T (B) field strengths. Only values within default mode network (DMN) ROIs are shown. On the right, boxplots represent the mean value of this difference inside the posterior cingulate cortex (PCC) ROI for all paired subjects. L‐TLE: left temporal lobe epilepsy. R‐TLE: right temporal lobe epilepsy. NS, not statistically significant.

3.5. rs‐fMRI lateralization ability at 7T is independent of MRI lesional status

We tested whether SOZ lateralization with 7T rs‐fMRI was contingent on the presence of structural abnormalities on standard‐of‐care clinical MRI (MRI lesional status). At 7T, the separation between L‐TLE and R‐TLE (Figure 6A) remained regardless of whether there was an obvious structural abnormality (lesional, Cohen's d = 2.94) or not (non‐lesional, Cohen's d = 1.83). At 3T (Figure 6B), a separation between L‐TLE and R‐TLE was evident in the lesional sub‐group (Cohen's d = 2.32), but not in the non‐lesional sub‐group (Cohen's d = 0.32). This might suggest that 3T rs‐fMRI may be able to lateralize TLE in lesional cases but not in non‐lesional cases, whereas 7T rs‐fMRI's lateralizing ability might be independent of lesional status. We also found that these findings are consistent in Engel I outcome subjects, which serves as a ground truth given their seizure freedom (Figure S7).

FIGURE 6.

Effects of lesional status, FDG‐PET hypometabolism, and sample size on lateralizing ability: (A, B) Show the mean difference between left hippocampal seed‐to‐voxel and right hippocampal seed‐to‐voxel maps within the PCC of subjects in the paired 7T and paired 3T cohorts, respectively. Each panel has subjects subdivided based on SOZ laterality and the presence (lesional) or absence (non‐lesional) of a structural abnormality on MRI. (C) Shows the quantitative FDG‐PET contrast asymmetry (right–left/right+left) for the (left) supplementary motor area, (middle) medial temporal lobe, and (right) temporal lobe across groups. (D, E) Show the correlation between the FDG‐PET temporal lobe asymmetry and the right–left hippocampus–PCC connectivity difference metric measured at (D) 7T and (E) 3T. (F, G) Show the same as (A) and (B), but for the extended 7T and 3T cohorts, respectively. Black circles represent subjects from the paired 3T–7T cohort. (H) Shows the receiver‐operating characteristic (ROC) curve for using the difference between the left hippocampal seed‐to‐voxel and right hippocampal seed‐to‐voxel maps within the PCC as a metric for classifying L‐TLE and R‐TLE in the extended 3T and 7T cohorts. The gray area represents the 95% confidence interval ROC generated by resampling the extended 3T cohort to the sample size of the extended 7T cohort. Similarly (I) (top) shows the Cohen's d distribution between L‐TLE and R‐TLE in (F) and (G), (bottom) as well as the area under the curve (AUC) from (H), after resampling the extended 3T cohort to the same sample size as the extended 7T cohort. The value of the extended 7T cohort for each of these metrics is shown as a vertical red line. TLE, temporal lobe epilepsy. PCC, posterior cingulate cortex; R‐TLE, right TLE; L‐TLE, left TLE; SOZ, seizure onset zone.

3.6. FDG‐PET temporal hypometabolism correlates with findings at 7T

Because FDG‐PET remains the neuroimaging clinical standard for SOZ lateralization in TLE, ³³ we compared quantitative FDG‐PET hypometabolism to our rs‐fMRI findings at 3T and 7T. We measured the asymmetry index of the FDG‐PET contrast within temporal and mesial‐temporal ROIs. As expected, FDG‐PET demonstrated temporal hypometabolism ipsilateral to the SOZ in both the left TLE (positive asymmetry) and right TLE (negative asymmetry), causing a statistically significant difference in asymmetry between groups (p _FWER = 0.009, Cohen's d = 1.90) (Figure 6C). No significant FDG‐PET asymmetries were detected in the SMA or in the mesial temporal lobe. We found a strong association between the FDG‐PET hypometabolism, as quantified by the asymmetry in FDG‐PET contrast, and the right–left hippocampus–PCC connectivity difference metric measured at 7T (Spearman ρ = 0.65, p = .01), but not at 3T (Spearman ρ = 0.13, p = .69) for the same subjects.

3.7. Effect of sample size on subject lateralization

Because small sample sizes can underestimate or overestimate measured effect sizes, we tested whether we could still lateralize the SOZ in the unpaired extended cohorts of TLE subjects scanned at 3T (n = 62) and 7T (n = 27). Findings in the extended cohorts replicated those in the original paired cohort, with the right–left hippocampus–PCC connectivity difference between L‐TLE and R‐TLE being statistically significant at 7T (p < .001, Cohen's d = 1.59) (Figure 6E). At 3T, we also saw statistically significant differences between L‐TLE and R‐TLE (p = .001, Cohen's d = 0.88) (Figure 6F), although the effect size was much lower than at 7T. Since the extended 3T cohort had more subjects than the extended 7T cohort, we also generated 1000 random subsets of the extended 3T cohort with the sample size of the extended 7T cohort. The Cohen's d in the extended 7T cohort (Cohen's d = 1.59) was higher than in 96% of the permuted equal sized subsets of the extended 3T cohort (Figure 6I). The ability to lateralize the SOZ using the right–left hippocampus–PCC connectivity difference as a metric, as with our paired 3T–7T cohort, was higher at 7T (ROC AUC 0.85, 95% CI: 0.66–1.00) for the extended 7T cohort, compared to at 3T (ROC AUC 0.73, 95% CI: 0.59–0.86). To ensure that these findings were not driven by biological and demographic variable differences between the extended cohorts, we performed a regression analysis controlling for these variables, which showed results consistent with those presented in this section (Tables S1 and S2; Figure S8).

4. DISCUSSION

In this study we examined the effect of magnetic field strength (3T vs 7T) on measuring resting‐state functional connectivity between the hippocampus and DMN nodes in patients with TLE. We found that although differences in connectivity patterns between L‐TLE and R‐TLE were in the same direction for 3T and 7T, the effect size was much larger at 7T for the same subset of subjects. We also demonstrated that the use of the hippocampal–PCC connectivity for lateralizing the SOZ showed superior performance at 7T compared to 3T. These results show that larger sample sizes at 3T may be able to replicate findings obtained at 7T but that 7T has better SOZ lateralizing ability in TLE regardless of sample size. Our findings also showed consistent results between the current neuroimaging standard FDG‐PET and 7T, but not 3T rs‐fMRI‐derived measurements.

The superior performance of 7T could be due to the increased subcortical tSNR in 7T acquisitions when compared to 3T. The higher subcortical tSNR at 7T could have been a driver for higher quality (less noisy) hippocampal seed‐to‐voxel maps, allowing for stronger measured effects in the PCC and MPFC. Because noise decreases as the sample size increases, by increasing the sample size of the 3T cohort we were able to better replicate the findings at 7T. Thus, 3T rs‐fMRI might still capture effects similar to those of 7T rs‐fMRI, but requires a much larger sample sizes.

Our results also corroborate prior findings regarding changes in hippocampal connectivity to the DMN in TLE. ⁸ , ⁹ , ¹⁰ , ¹¹ , ¹² , ¹³ However, much of this past work has focused on group level analyses, making it challenging to assess the potential for clinical translation. We provide evidence that DMN connectivity findings in TLE can be replicated, and potentially expanded, at 7T, allowing for stronger effect sizes at a group level, and increased clinical utility at a single subject level. Despite the widespread availability of fMRI and convergent evidence showing its potential in lateralizing epilepsy, the integration of fMRI into the presurgical pipeline remains limited. We maintain that the lateralized changes we observed in the PCC for left and right TLE support the applicability of fMRI in pre‐surgically lateralizing epilepsy. Furthermore, the superior performance of 7T in lateralizing TLE could provide an avenue for fMRI to be more widely used during presurgical evaluation for seizure focus lateralization purposes, thereby complementing the findings of other functional imaging modalities such as MEG, FDG‐PET, and single photon emission computer tomography (SPECT).

The main limitation of our study is the sample size of our paired 3T–7T cohort. Increasing the sample size of the paired cohort would allow better understanding of the differences between 3T and 7T fMRI acquisitions in epilepsy and may allow a better exploration of regions/networks of interest other than the DMN. Although we added more subjects to the 3T‐only and 7T‐only cohorts, these groups inevitably drive the extended analyses; biological differences in the distribution of network abnormalities across these separate cohorts cannot be fully disentangled from the differences related to field strength. Our regression analyses (including age, sex, and presence of MTS) partly address this concern; however, they do not entirely eliminate the possibility of confounding effects.

Beyond cohort‐related factors, 7T imaging itself poses unique challenges. Because 7T magnets remain comparatively scarce, patients often must travel, which raises costs and logistical barriers. The scanners typically have smaller but longer bores and bulkier head coils, creating added difficulties for individuals with claustrophobia or larger body habitus. From a technical standpoint, 7T images are more prone to pronounced susceptibility artifacts in basal temporal regions, have steeper signal drops in inferior brain areas, and exhibit stronger B1 bias. These effects require more sophisticated post‐processing and must be taken into consideration when weighing the potential benefits of 7T for SOZ lateralization.

Future studies should focus on acquiring a larger paired 3T–7T dataset as well as multi‐institutional datasets that could help further demonstrate the utility of 7T rs‐fMRI. Additionally, although our study focused on changes within a defined brain natwork, the DMN, and specifically the PCC and MPFC, some of our findings (the superior subcortical tSNR at 7T, particularly in the thalamus), might suggest that there is an avenue to further explore the subcortical connectivity profile of epilepsy using ultra‐high field imaging.

5. CONCLUSION

Our study provides a direct comparison between 3T and 7T rs‐fMRI in lateralizing TLE. Our findings show that although connectivity patterns between L‐TLE and R‐TLE were in the same direction for 3T and 7T, the effect size was much larger at 7T for the same subset of subjects, and the use of the hippocampal–PCC connectivity for lateralizing the SOZ showed superior performance at 7T compared to 3T. This supports the adoption of ultra‐high field functional neuroimaging in the presurgical assessment of epilepsy.

AUTHOR CONTRIBUTIONS

Alfredo Lucas: Conceptualization, methodology, software, validation, formal analysis, and writing—original draft preparation. Eli J. Cornblath: Conceptualization, writing—review & editing, and data curation. Nishant Sinha: Conceptualization and writing—review & editing. Lorenzo Caciagli: Conceptualization and writing—review & editing. Peter Hadar: Data curation and writing—review & editing. Ashley Tranquille: Data curation and writing—review & editing. Joel M. Stein: Data curation, conceptualization, and writing—review & editing. Sandhitsu Das: Data curation, conceptualization, and writing—review & editing. Kathryn A. Davis: Supervision, project administration, funding acquisition, data curation, conceptualization, and writing—review & editing.

CONFLICT OF INTEREST STATEMENT

The authors report no competing or financial interests.

ETHICS STATEMENT

We confirm that we have read the Journal's position on issues involved in ethical publication and affirm that this report is consistent with those guidelines.

Supporting information

Data S1.

Lucas A, Cornblath EJ, Sinha N, Caciagli L, Hadar P, Tranquille A, et al. Seizure‐onset zone lateralization in temporal lobe epilepsy using 7T rs‐fMRI: Direct comparison with 3T rs‐fMRI . Epilepsia. 2025;66:3440–3452. 10.1111/epi.18447

DATA AVAILABILITY STATEMENT

The code and preprocessed data to reproduce the analyses and figures for this manuscript is available here: https://github.com/penn‐cnt/3t_7t.