Source Reconstruction

Individual forward models were computed using three-layer boundary element models (BEMs) based on FreeSurfer cortical reconstructions (Fischl, 2012), with source spaces at 4,098 vertices per hemisphere (ico4 spacing). Session-specific noise covariance was estimated from 400 ms pre-scene baseline periods, pooled across all 10 sessions, using Oracle Approximating Shrinkage (OAS). The inverse solution used loose orientation constraint 0.2, depth weighting 0.8, and an assumed SNR of 3 (lambda-squared = 1/9). Source estimates were computed with dSPM (Dale et al., 2000) as implemented in MNE-Python, and morphed to the fsaverage template (ico5 spacing, smoothing = 5).

Category-averaged event-related fields (ERFs) were averaged over a +/-10 ms window centered on the group-level sensor-space representational-similarity peak latency (114 ms post-fixation onset – see Example Analyses). A geodesic searchlight (20 mm radius) was applied across the cortical surface using MNE-RSA (van Vliet et al., 2025). Visual regions of interest (early, lateral, ventral, parietal) were defined following the NSD cortical ROI scheme (Allen et al., 2022); frontal regions (dlPFC, FEF, OFC, mPFC, infFC) and hippocampus were defined as compound regions from the Glasser parcellation (Glasser et al., 2016).

Running This With pyAVS

pyAVS’s source module (pyavs.source, see Source Reconstruction (pyavs.source)) implements this pipeline: pyavs.create_bem_model() / pyavs.create_source_space() / pyavs.setup_coregistration() / pyavs.load_forward_model() for forward modeling, pyavs.compute_beamformer_filters() and pyavs.apply_source_reconstruction() for beamformer-based reconstruction, and pyavs.get_roi_labels() / pyavs.get_glasser_roi_labels() / pyavs.extract_roi_data() for ROI-level summarization. See Source Reconstruction and Population Codes for a guided walkthrough and Source Reconstruction Examples for runnable code.

Note

pyAVS’s example pipelines use LCMV beamforming rather than dSPM by default (both are supported); the dSPM/BEM/fsaverage configuration described above is specifically the one used for the source-projected fixation ERFs reported in the dataset paper.