Evaluating and optimizingfMRI processing pipelines with the NPAIRS approach for decision support in fMRI applications

This study explores evaluation and optimization of BOLD fMRI processing pipelines with a cross-validation, resampling approach called NPAIRS. The original NPAIRS package can not incorporate heterogeneous software modules nor generate GLM (General Linear Model) prediction metrics. To overcome these limitations, a Java-based pipeline-evaluation system has been developed that applies NPAIRS performance metrics (prediction, reproducibility) to GLM and Canonical Variates Analysis (CVA) models, and that uses modules from heterogeneous fMRI data-processing packages.; In this study, the impact of commonly-used preprocessing steps has been investigated, associated single-subject processing pipelines have been optimized, and pipelines with GLM and CVA-based data analysis models from multiple software packages have been evaluated.; For block-design fMRI data from a parametric, static-force task, it is found that, (1) slice timing correction and global intensity normalization have little consistent impact on fMRI processing pipelines, but spatial smoothing and high-pass filtering or temporal detrending significantly increases, while prewhitening significantly decreases pipeline performance; (2) combined optimization of spatial smoothing and temporal detrending preprocessing steps improve pipeline performance and on average improve between subject reproducibility; (3) among the fMRI processing pipelines evaluated, the most important choices include spatial smoothing and univariate-or-multivariate statistical models; (4) In general, the prediction performance of CVA models is higher than that of the GLM models, while GLM models have higher reproducibility performance than CVA models.; This study proves that: (1) it is feasible to overcome the limitations of the current NPAIRS package and apply NPAIRS metrics to new analytic modules by connecting the Java-based Fiswidgets fMRI software environment to the YALE machine-learning environment; (2) the NPAIRS approach may be used to systematically rank the importance of pipeline choices; (3) multivariate CVA models are more sensitive to noise and preprocessing choices than univariate GLM models; (4) because of the bias-variance trade-off of the GLM and CVA models, it may be necessary to consider a consensus approach for better decision support in fMRI-aided applications; (5) Considering non-standard options other than the widely-used GLM with a fixed spatial filter may be of critical importance in determining reliable activation patterns in BOLD fMRI studies.