Predicting cortical signal mixing in coarse-resolution fMRI: a forward-operator analysis of surface mapping and the conditional benefit of correction

Abstract

Surface-based fMRI projection can mix signals from cortical locations that are close in the volume but distant along the folded cortical sheet. We constructed an anatomy- and acquisition-informed linear forward operator to estimate this mixing and evaluated a regularised surface reconstruction that inverts it. In HCP 7T geometry, the operator-predicted net far-cortex contribution to a standard surface value exceeded 5% in 4%, 10% and 22% of the group-average cortex at 1.6 mm and at emulated 2.4 and 3.2 mm resolution. Including the modelled far component improved predictions of emulated coarser maps and of real coarser StudyForrest sessions in high-leakage regions. Against finer-resolution, cross-run fMRI references, the regularised reconstruction reduced the residual error in HCP development data (n = 15) and in a pre-specified, family-independent confirmatory analysis (158 participants in 86 families). The improvement increased with a response-weighted contamination measure in HCP and in an independent 3T dataset (35 test participants), although a coarse-data-only version of the measure did not generalise to 3T, where an anatomical atlas predicted the gain at twofold coarsening. Overall performance improved at twofold, but not 1.5-fold, coarsening in the 3T sample. Ablation and data-reduction analyses were consistent with an important contribution from local deconvolution but did not uniquely identify the mechanism. Benefits were smaller or absent for multi-run averages, and no robust improvement across separately acquired sessions was established. The leakage maps characterise model-predicted, signed mixing rather than directly measured signal origins, and every gain is measured against an fMRI reference rather than against the true cortical response. Code, analysis plans and group leakage maps (7T and 3T) are provided in a public repository.