Analysis Methods¶
Microstructure profile covariance gradients¶
Microstructural profile covariance (MPC) was estimated by sampling intracortical signal intensities from a microstructure-sensitive quantitative T1 (qT1) contrast. Fourteen equivolumetric cortical surfaces were generated between the pial and white matter boundaries, enabling the extraction of vertex-wise intracortical intensity profiles along the cortical ribbon. The Salience Network (SN) was defined according to the Yeo seven-network parcellation (Schaefer-400), corresponding to the ventral attention network (VAN). For each subject, MPC matrices were constructed by computing partial correlations between vertex-wise profiles while controlling for the mean cortical profile as a covariate, thereby capturing the similarity of intracortical microstructural organization across vertices within the salience network. Low-dimensional representations (gradients) of salience network MPC were derived using the BrainSpace toolbox. Ten gradient components were extracted using diffusion map embedding, a nonlinear manifold learning approach based on the graph Laplacian. Affinity matrices were constructed using a normalized angle kernel with a sparsity threshold of 0.9, and resulting gradients were aligned across subjects using Procrustes rotation.
Cortical type reconstruction¶
Cortical types were assigned to Von Economo areas based on a recent reanalysis of Von Economo micrographs. This classification scheme was used because its criteria are (1) clearly defined, (2) applied consistently across the entire cortex, (3) align with Von Economo's original descriptions and (4) are supported by several histological samples. Criteria included 'development of layer IV, prominence (denser cellularity and larger neurons) of deep (V–VI) or superficial (II–III) layers, definition of sublayers (for example, IIIa and IIIb), sharpness of boundaries between layers and presence of large pyramids in superficial layers'. Thereby, cortical types synopsize degree of granularity, from high laminar elaboration in koniocortical areas, six identifiable layers in Eu-III to -I, poorly differentiated layers in dysgranular and absent layers in agranular.
Structural network reconstruction¶
DWI pre-processing was implemented with micapipe DWI module, which heavily relies on tools from MRtrix. More specifically, fiber orientation distributions were generated using the multi-shell, multi-tissue constrained spherical deconvolution (msmt-CSD) algorithm from MRtrix. 40 million streamlines were then reconstructed using an anatomically constrained probabilistic tractography algorithm (ACT-iFOD2) based on the generated fiber orientation distributions (fODF). The connectivity weights were then optimized using the Spherical-deconvolution Informed Filtering of Tractograms (SIFT2) by estimating an appropriate cross-section multiplier for each streamline based on apparent fiber density (AFD), and a connectivity matrix was built for each participant using the 400-region Schaefer parcellation. Associated edge length matrix consisting of the mean streamline length between each node pair. Communication in the structural connectome was modeled using navigation implemented in bctpy toolkit. Navigation combines the structural connectome with euclidian distance between parcellation nodes. In brief, navigation involves identifying a single, efficient path between two nodes, where each step is determined by spatial proximity to the target node. Specifically, the next node in the path is the neighbor of the current node (that is, sharing a structural connection) that is closest to the final target node. Navigation is the sum of connection weights as defined by SC along the navigation path. Navigation was calculated within each hemisphere separately, then concatenated for analyses.
iEEG signal processing¶
iEEG signals (MNI open iEEG atlas and MICA iEEG datasets) were preprocessed using a common pipeline. Raw signals were band-pass filtered between 0.5 and 80 Hz using a 4th-order zero-phase Butterworth filter, then downsampled to 200 Hz. Signals were subsequently demeaned by subtracting the temporal mean of each channel. Power spectral density (PSD) was estimated using Welch's method with a Hamming window of 2-second segments and 1-second overlap. Band power was computed by integrating the PSD within each canonical frequency band using Simpson's rule. Relative band power was obtained by dividing each band's integral by the total power, and log₁₀-transformed (floor = 1×10⁻¹²). Frequency bands were defined as: delta (0.5–4 Hz), theta (4–8 Hz), alpha (8–13 Hz), beta (13–30 Hz), and gamma (30–80 Hz).
Brain surface feature comparisons¶
Whole brain surface feature comparison was implemented with brainspace using sphere spin permutation (23) with 100 permutations to generate null data for hypothesis testing. Salience network brain surface feature comparison was implemented with brainspace using moran spectra randomisation (24) with 100 permutations, which uses the eigenvectors to generate null model data with similar spatial autocorrelation. The implemented procedure "singleton" matches the input data's autocorrelation more closely at the cost of fewer possible randomizations.