Cross-modal quality transfer: enhancing MEG spatial resolution using BOLD-fMRI and Explainable machine learning.

Purpose: Blood-oxygen-level-dependent functional MRI (BOLD-fMRI) and magnetoencephalography (MEG) offer complementary insights into brain function, with BOLD-fMRI providing high spatial resolution and MEG offering high temporal sensitivity. This study aimed to enhance MEG spatial resolution through learning inter-modal relationships via data-driven fusion with BOLD-fMRI using explainable machine learning (xML). Methods: MEG and BOLD-fMRI data were collected for sixteen participants watching the same naturalistic visual stimulus. MEG signals were filtered into standard frequency bands and processed to match the haemodynamic response. Data were then decomposed using Tensorial Independent Component Analysis (TICA), yielding 250 MEG (25 per frequency band) and 30 BOLD-fMRI components. A set of Extreme Gradient Boosted (XGBoost) models were trained to predict MEG component activity from downsampled BOLD-fMRI components. The models with best-performing configuration of hyperparameters at this resolution were then used to generate voxelwise upsampled MEG maps using the natively higher-resolution BOLD-fMRI data. Performance was compared to trilinear interpolation using R2, mean squared error (MSE), and structural similarity index (SSI). Model interpretability was enhanced by generating Shapley values (SHAP), describing relationships between input data and model output. Results Models trained on 6mm3 data achieved high predictive performance on previously unseen data (R2 = 0.80, MSE = .07). Using higher-resolution 2mm3 BOLD-fMRI data, MEG activity was upsampled to 2mm3, resulting in more detailed image, while maintaining reasonable congruence to naive interpolation (R2 = .81, MSE = .06). Conclusions: This work demonstrates that explainable machine learning enables spatial super-resolution of MEG maps via BOLD-fMRI-informed quality transfer, enabling enhancement of spatial resolution, and offering improved localization, as well as means of generating data-driven insights into neurovascular coupling.