Spatially adaptive subject level analyses improve random effects fMRI group studies

Abstract

Inter-subject analysis of functional Magnetic Resonance Imaging (fMRI) data relies on single intra-subject studies, which are usually conducted using a massively univariate approach. In this paper, we investigate the impact of an improved intra-subject analysis on group studies. Our approach is based on the use of Adaptive Spatial Mixture Models within a joint detection-estimation (JDE) framework [1]. In this setting, spatial variability is achieved at a regional scale by the explicit characterization of the hemodynamic filter and at the voxel scale by an adaptive spatial correlation model between condition-specific effects. For the group statistics, we conducted several Random effect analyses (RFX) which relied either on SPM or JDE intra-subject analyses. We performed a comparative study on two different real datasets involving the same paradigm and the same 15 subjects but eliciting different noise levels by varying the acceleration factor (R=2 and R=4) in parallel MRI acquisition. We show that brain activations appear more spatially resolved using JDE instead of SPM and that a better sensitivity is achieved. Moreover, the JDE framework provides more robust detection performance by maintaining satisfying results on our most noisy real dataset.