Scalable and Adaptive Spatiotemporal Modeling for Task-Based fMRI Analysis

Task-based fMRI is commonly analyzed using voxel-wise general linear models, a non-spatial scalable approach that can yield fragmented activation maps. Spatial alternatives such as kernel smoothing and Bayesian models address this but either blur activation boundaries or are computationally prohibitive at modern spatial resolutions. We introduce SPLASH (Spline-Based Processing for Localized Adaptive Spatial Hemodynamics), a spatially adaptive and scalable framework based on localized thin-plate spline regression within brain parcels. Its spatial flexibility allows SPLASH to adapt to heterogeneous cortical organization and to generalize across diverse spatial domains. Using its hierarchical structure, we introduce a two-stage selective inference procedure that ensures valid false discovery rate control at the parcel and voxel levels. In simulations, SPLASH consistently delivered the best overall performance: its MSE was typically only 20-40% of that of prior spatial models, and both FPR and FNR remained well controlled. SPLASH also remained stable across smoothing choices and required only 2% of the computation time of Bayesian spatial approaches. Applied to Human Connectome Project data, SPLASH produced sharper activation patterns consistent with the motor homunculus and demonstrated higher reproducibility. SPLASH provides a generalizable, spatially adaptive, and scalable framework that strengthens statistical inference and improves neuroscientific interpretability in large-scale fMRI studies.