Cortical parcellation optimized for magnetoencephalography with a clustering technique

A typical approach to estimate connectivity from magnetoencephalographic (MEG) data consists of 1) computing a cortically-constrained, distributed source estimate, 2) dividing the cortex into parcels according to an anatomical atlas, 3) combining the source time courses within each parcel, and 4) computing a connectivity metric between these combined time courses. However, combining MEG signals to spatial mean activities of anatomically-defined parcels often leads to cancellation within and crosstalk between parcels. We present a method that divides the cortex into parcels whose activity can be faithfully represented by a single dipolar source while minimizing inter-parcel crosstalk. The method relies on unsupervised clustering of the MEG leadfields, also accounting for distances between the cortically-constrained sources to promote spatially contiguous parcels. The cluster each source point belongs to is determined by its k nearest-neighbour memberships. Inter-parcel crosstalk was minimized by assigning and a weight of 20%-40% to the spatial distances, leading to 60–120 parcels. Our approach, available through the Python package “megicparc”, enables a compact yet anatomically-informed source-level representation of MEG data with a similar dimensionality as in the original sensor-level data. Such representation should enable significant improvements in source-space visualization of MEG features or in estimating functional connectivity.