CiftiStorm pipeline: facilitating reproducible EEG/MEG source connectomics

Abstract

We present CiftiStorm, an electrophysiological source imaging (ESI) pipeline incorporating recently developed methods to improve forward and inverse solutions. The CiftiStorm pipeline produces Human Connectome Project (HCP) and megconnectome-compliant outputs from dataset inputs with varying degrees of spatial resolution. The input data can range from low-sensor-density electroencephalogram (EEG) or magnetoencephalogram (MEG) recordings without structural magnetic resonance imaging (sMRI) to high-density EEG/MEG recordings with an HCP multimodal sMRI compliant protocol. CiftiStorm introduces a numerical quality control of the lead field and geometrical corrections to the head and source models for forward modeling. For the inverse modeling, we present a Bayesian estimation of the cross-spectrum of sources based on multiple priors. We facilitate ESI in the T1w/FSAverage32k high-resolution space obtained from individual sMRI. We validate this feature by comparing CiftiStorm outputs for EEG and MRI data from the Cuban Human Brain Mapping Project (CHBMP) acquired with technologies a decade before the HCP MEG and MRI standardized dataset.

Keywords: Brainstorm; Ciftify; HIGGS; SSSBL; VARETA; forward model; human connectome project; megconnectome.

Figure 1

The conceptual organization of CiftiStorm. Red boxes highlight the conceptually different pipelines, and yellow boxes highlight the pipeline modules within CiftiStorm. In the top row is the definition for Brainstorm inputs within a dataset, which may include individual T1w, T2w, or none of these sMRI acquisitions; the MNI-registered sMRI template that better describes the dataset; and the EEG or MEG acquisitions. The modules include the standard and legacy modules of an HCP-compatible structural processing pipeline on the second row, the source model, head model, and lead field modules of a forward model processing pipeline on the third row, and the prior’s integration and inverse solution modules of an inverse model processing pipeline at the fourth row. CiftiStorm incorporates quality control and corrections (blue looped arrows) driven by stringent factors derived from processing by the different pipeline modules. Within the scope of our study is the production of highly similar EEG/MEG inverse solutions, illustrated in the green box, ensuring broader EEG ESI integration.

Figure 2

Illustration of the CiftiStorm structural processing pipeline. Top row: (A) the HCP standard PreFreeSurfer, FreeSurfer, and PostFreeSurfer pipeline modules that are applied to T1/T2w sMRI, (B) the HCP legacy FreeSurfer and Ciftify pipeline modules that are applied to T1w or sMRI (legacy databases), and (C) HCP spaces (volumetric and superficial) in the T1w (MNILinear) or MNINonlinear, and Native or FSAverage32k. Bottom row: (D) cortical curvature, (E) thickness, (F) parcellation, (G) volumetric subcortical segmentation, and (H) cortical layers mid-thickness and pial.

Figure 3

Megconnectome or CiftiStorm forward and inverse model processing leverage the surface-based source model and metrics. (A) The geometry includes triangular meshes (top), curvature or second-order surface metric (middle), and normal or first-order surface metric projected in the Z-axis (bottom). (B) Surface-based geometrical information, including the discrete Laplacian prior (top), curvature depth compensation prior (middle), and rotational invariance prior (bottom), are to be considered by inverse modeling methods. (C) Shows the prior representation included in our strategy for the source processing module, Graph Laplacian (top), curvature depth compensation prior (middle), and rotational invariance (bottom). In addition, a typical inverse solution was registered across all HCP spaces and interpolated to high resolution by the CiftiStorm pipeline. The inverse solution is shown in (D) the FSAverage32k canonical space, (E) the T1w/Native space, and (F) the MNI-ICBM152 space.

Figure 4

CiftiStorm head model, lead field processing, and quality control workflows. Top box: The neuroinformatic tools included in our pipeline and their different roles. Bottom box: The primary outputs are later targeted by the manual or automatic quality control analysis. An illustration aims to show the expected outputs under ideal working conditions, obtained from the high-quality MNI-ICBM152 image and its legacy structural processing (Sec. 2.1.2) and source model processing (Sec. 2.2.1). Outputs are (A) the head model comprising four structures: cortex, inner skull, outer skull, and scalp surfaces. These surfaces are due to the HCP brain and FSL head processing, postprocessing, and registration via the Brainstorm head modeler utility. (B) The registration for EEG electrodes and their layout with the scalp surface. (C) The registration with the scalp surface for MEG magnetometers and their helmet layout. For optimal sensor registration in (B) and (C), the CiftiStorm pipeline leverages utilities from both the Brainstorm and the SPM head modelers. (D) The lead field for an EEG electrode. (E) The lead field for a MEG magnetometer. We use the reciprocal representation of electromagnetism to illustrate at the source space (cortex) the distinct vector field regimens for EEG and MEG sensors. Both are due to BEM computations employing the Brainstorm OpenMEEG or alternative FieldTrip utilities.

Figure 5

Quality control indicator information of the same case is shown in Figure 4 in an expanded way, which is also included in CiftiStorm’s graphic interface and printed reports. These indicators include (A) the sMRI and the registration of source and head models with it (from left to right: the scalp, outer skull, inner skull, and brain surfaces); (B) mutual registration between the head model and the source model, from left to right, first one: the scalp, outer skull, inner skull, and brain surfaces, second one: inner skull, outer skull and brain cortex, third one: inner skull and brain cortex, last one outer skull and inner skull; (C) left, right and top view of the inner skull and brain cortes registration, in red color: the points situated too close between them; (D) inner skull and brain cortex registration after distance correction process; (E) the vector field distribution (blue arrows) for a user selected sensor (red dot), sensors (green dots); (F) correlation tests between the tested lead field and the homogeneous lead field and their corresponding Pearson correlation coefficient, first one: tested and homogeneous lead field correlation, second one: sensor-wise correlation between tested and homogeneous lead field, and last on: source-wise correlation between tested and homogeneous lead field.

Figure 6

Illustration of the different notions embodying the cross-spectral methodology VARETA. (A) Characteristic spectral sensor topographies and cortical topographies composing the human resting state EEG spectrum. The cortical topographies that are determined by a conventional or cross-spectral inverse solution are the brain signature for band-limited oscillatory networks and the mechanism for activity associated with particular cognition and behavior. (B) Cross-spectral analyses of EEG sensor data and their sources involved in the cross-spectral inverse solution device, which follows the Fourier transform path. We describe this path as a series of tensor variables. The targeted variable is the theoretical source cross-spectral tensor, which must be determined by different methods from its sampled equivalent calculated for the data or the sources. (C) Regularization modes to resolve the high spatial ill-condition and dimensionality in the tensor target. These modes target independent distortions, measured as type-one and type-two leakage corresponding to the spectra and cross-spectra.

Figure 7

Case flagged as “artifactual” for two major geometrical artifacts in Figure 5 and their solution in the quality control loop. (A) Individual sMRI showing movement artifact and the corresponding artifactual source model and head model outputs. (B) ICBM-152 sMRI, source model, and head model template substituting the individual outputs. (C) Incorrect EEG layout. (D) Corrected EEG layout.

Figure 8

Case flagged as “artifactual” for the lead field artifacts in Figure 5 and their solution in the quality control loop. We show the minimum correlation value between the homogeneous and actual lead field outputs before (red) and after (green) correction in (A) across sources (for each sensor) and in (B) across sensors (for each source). The linear fit of both lead fields relative to the same correction is shown in (C) before and (D) after. (E) the artifactual, homogeneous, and corrected lead fields are shown from left to right. (F) From left to right, and under distance criteria, the sources labeled (red) as too close to the inner skull surface and the sources after correction. (G) From left to right, there is a cortical colormap of the source correlations before and after correction.

Figure 9

The CiftiStorm inverse solution pipelines from the CHBMP EEG and HCP MEG datasets show differences between their source spectra. (A) Scattergram manifold illustrating the global differences between MEG and EEG source $\log$ spectra. The large differences are due to the properties of either type of acquisition or preprocessing style. (B) Results of the regression using a linear model that represents these differences as a scale and a slope parameter. (C) Goodness of fit for this linear model. (D) Random permutation test outcomes of the differences for delta, theta, alpha, beta, and gamma-low bands.