Clinical epilepsy assessment involves localizing seizure onset zones (SOZs) and networks (SNs) non-invasively based on extra-cranial EEG and identifying focal slowing, high-frequency oscillations, and various epileptiform discharges^1-3. However, EEG has moderate diagnostic sensitivity (40%) and poor spatial specificity (1-5 cm)⁴. Towards more effective non-invasive characterization of SOZs and SNs than currently established, we implemented multi-echo multi-band (ME-MB) BOLD fMRI at 7T to map seizure-related networks at millimeter-resolution and characterize epileptiform hemodynamics across BOLD frequencies.

Three patients with focal non-lesional (1 female, mean age 26y) and two with focal lesional epilepsy (1 female, mean age 24y) and matched controls participated in this 7T MRI study, which was approved by the Mt. Sinai IRB. Lesion diagnosis was by a board certified and experienced radiologist, and epilepsy diagnosis was by certified epileptologists at the Mt. Sinai Epilepsy Center. All participants were scanned on a Siemens Magnetom 7T MRI scanner (Siemens, Erlangen, Germany) with a birdcage-transmit/32-channel-receive head coil. Scanning sequences included MP2RAGE⁵ (0.8mm iso., TR/TE=6s/5ms, TI=1050ms;3000ms), T2-FLAIR (0.7x0.7x3mm, TR/TE/TI=9000/123/2600ms), and 8 minutes of "resting state" MEMB-fMRI (whole-brain 2.5mm iso.; TR=1.85s, TEs=8.5,23.2,37.8, 52.5ms; MB=3; GRAPPA=3)^6-7. MP2RAGE “UNI” images homogenized T1 contrast and attenuated intensity non-uniformity, and were used for brain extraction and masking T2-FLAIR images co-registered with T1 images. MEMB-fMRI analysis involved multi-echo independent components analysis (ME-ICA) with AFNI meica.py⁸ with default settings, which implemented: slice time and motion correction; co-registration to masked T1; T2* mapping and weighted "optimal combination" of echoes⁹; dimensionality estimation with multi-echo PCA (ΔT2* fitting of PC amplitude vs. TE); spatial FastICA decomposition; BOLD/non-BOLD IC selection (ΔT2* fitting of IC amplitude vs. TE), and time series denoising by non-BOLD component removal – without arbitrary noise models for head and cardiopulmonary motion, temporal bandpass filtering, spatial smoothing, etc.^10-12 BOLD IC (i.e. functional network) time series were compared of patients and controls to determine differences in 1) frequency spectra and 2) statistical moments reflecting sparsity to infer “bursting.” Power spectra were computed for all BOLD component time courses (temporally Z-normalized), collapsed within control and patient groups, and then amplitudes per frequency bin were compared between patients and controls using a 2 independent-samples T-test. Sparsity was assessed for each IC time course as skewness (σ₃) and kurtosis (σ₄), and groups were compared with Mann-Whitney U-tests. Individual functional network IC maps were manually examined for the lesional epilepsy patients, respectively with hippocampal cavernoma (venous abnormality) and temporal lobe dysembryonic neuroepithelial tumor (DNET, development abnormality). Networks were examined to elucidate if BOLD networks would co-localize with anatomical abnormalities (putative SOZs) and suggest SNs.7T MEMB time series after T2* weighted combination compensated orbitofrontal susceptibility artifact without specialized shim/hardware (Figure 1a, patient). The default mode network was found in all patients and controls (Figure 1b). BOLD network ICs totaled 165 and 115 for [all] patients and matched controls, respectively. BOLD network ICs of patients had significantly higher spectral power in f<0.01 and f>0.1 Hz frequency ranges versus controls (p<10^-5, p<10^-10, respectively, Figure 2a-b). Notably, the average difference of BOLD spectral power in the canonical 0.01-0.1 Hz range was not significant. Additionally, network IC time course sparsity was significantly higher for patients than controls, in both skewness and kurtosis (both p<10^-6; Figure 2c-d), suggesting greater hemodynamic bursting activity in patients. Patient functional network maps showed: 1) hippocampal cavernoma associated with a unilateral hippocampal-temporal lobe functional network not resembling a canonical¹³ network (Figure 3a) but consistent with clinically determined seizure network and showing non-stationary time course transitions between high-frequency/low-amplitude and low-frequency/high-amplitude states; 2) DNET (dark in T1, bright in T2) was associated with a bilateral auditory cortex network (Figure 3b), notable since patient heard buzzing during seizure aura.7T MEMB-fMRI with ME-ICA processing elucidated patient functional networks at millimeter-resolution and mitigated artifacts from magnetic susceptibility as well as subject motion, in a comprehensive but unbiased way, critically reducing processing-related analysis confounds. This approach specially supported using MB-fMRI at 7T to find SNs as non-canonical functional networks co-localized with lesion/SOZs and observing temporal BOLD hemodynamics of significantly higher infra-slow (<0.01Hz) and high-frequency (>0.1Hz) amplitude and bursting in patients than controls. With future development and application, these techniques may lead to rapid high-resolution SN and SOZ mapping based on MRI, more effectively and earlier in medical or surgical treatment planning than currently achievable with EEG.Our 7T MEMB-fMRI approach is promising towards achieving high sensitivity and specificity for ultra-high-field functional imaging of epilepsy and other neuropsychiatric patients - at individual level - with the multi-echo component enabling identification of networks localized to abnormal tissue and/or exhibiting aberrant hemodynamics potentially coupled to pathological electrophysiology.