The Impact of Spatial Normalization for Functional Magnetic Resonance Imaging Data Analyses Revisited

Spatial normalization--the process of aligning anatomical or functional data acquired from different individuals to a common stereotaxic atlas--is routinely used in the vast majority of functional neuroimaging studies, with important consequences for scientific inference and reproducibility. Although several approaches exist, multi-step techniques that leverage the superior contrast and spatial resolution afforded by T1-weighted anatomical images to normalize echo planar imaging (EPI) functional data acquired from the same individuals (T1EPI) is now standard. Yet, recent work suggests that direct alignment of functional data to a T2*-weighted template without recourse to an anatomical image--an EPI only (EPIO) approach--enhances normalization precision. This counterintuitive claim is intriguing, suggesting that a change in standard practices may be warranted. Here, we re-visit these conclusions, extending prior work to encompass newly developed measures of normalization precision, accuracy, and real-world statistical performance for the standard EPIO and T1EPI pipelines implemented in SPM12, a recently developed variant of the EPIO pipeline, and a novel T1EPI pipeline incorporating best practice tools from multiple software packages. The multi-tool T1EPI pipeline was consistently the most precise, most accurate, and resulted in the largest t values at the group level, in some cases dramatically so. The three SPM-based pipelines exhibited more modest and variable differences in performance relative to each another, with the widely used T1EPI pipeline showing the second best overall precision and accuracy, and the recently developed EPIO pipeline generally showing the poorest overall performance. The results demonstrate that standard pipelines can be easily improved and we encourage researchers to invest the resources necessary to do so. The multi-tool pipeline presented here provides a framework for doing so. In addition, the novel performance metrics described here should prove useful for reporting and validating future methods for pre-processing functional neuroimaging data.