Exploring fMRI Results Space: 31 Variants of an fMRI Analysis in AFNI, FSL, and SPM

摘要

Background Data sharing is becoming a priority in functional Magnetic Resonance Imaging (fMRI) research, but the lack of a standard format for shared data is an obstacle ( Poline et al., 2012 ; Poldrack and Gorgolewski, 2014 ). This is especially true for information about data provenance, including auxiliary information such as participant characteristics and task descriptions. The three most commonly used analysis software packages [AFNI~( 1 )( Cox, 1996 ), FSL~( 2 )( Jenkinson et al., 2012 ), and SPM~( 3 )( Penny et al., 2011 )] broadly conduct the same analysis, but differ in how fundamental concepts are described, and have a myriad of differences in the pre-processing and modeling steps. The practical consequence is that sharing analyzed data is further complicated by the idiosyncrasies of the particular software used. The Neuroimaging Data Model [NIDM~( 4 )( Keator et al., 2013 ; Maumet et al., 2016 )] is an initiative from the International Neuroinformatics Coordinating Facility (INCF~( 5 )) that addresses these practical barriers through the development of a standard format for neuroimaging data. Ultimately, NIDM will provide a standard format that can handle data that has been processed in any of the common software packages. In order to achieve this, the development of NIDM requires publicly available derived data that covers all the major use cases in the main software programs. The purpose of the current work was to produce a set of results of mass univariate fMRI analyses using the most common software packages: AFNI, FSL, and SPM [which between them cover 80% of published fMRI analyses ( Carp, 2012 )], utilizing publicly available data from OpenfMRI~( 6 )( Poldrack et al., 2013 ). The analyses (‘variants’) presented in this paper cover the most common options available in each software package at each analysis stage, from different Hemodynamic response function (HRF) basis functions through to group-level tests. The tests are arranged so that readers can compare the closest equivalent variants across software packages. In particular, these tests will be useful for comparing the results from default test settings across software packages. While this collection of analyses was chosen for their relevance to the NIDM project, it also addresses a gap in the literature where publicly available processed data is concerned. Specifically, while there are published comparisons of different processing pipelines, the data are not publicly available ( Carp, 2012 ) or are for resting state fMRI only ( Bellec et al., 2016 ). Others have shared raw data but lack analysis results ( Hanke et al., 2014 ) or do not include comparisons across multiple software packages [e.g., analyses in the The Human Connectome Project~( 7 )( Van Essen et al., 2013 ) are performed with FSL only]. Shared raw data is a useful resource, but we argue that shared processed data is also important, both to provide a basis of cross-software comparisons and to create a benchmark for testing of automated provenance software. The dataset presented in this paper is a contribution toward this omission in the literature. Methods Data Source Data were downloaded from OpenfMRI’s BIDS-compliant ds000011 dataset~( 8 )between 09/02/2016 and 15/02/2016. A full description of the paradigm is in the original paper ( Foerde et al., 2006 ). The first task was a training exercise in which participants counted high tones in a series of high-pitched and low-pitched tones (‘tone counting’ condition), and then selected a number that represented the number of high tones (the ‘tone counting probe’ condition, referred to as ‘probe’ hereafter). We modeled both the tone counting and probe conditions, using tone counting as the effect of interest ([1 0] contrast with implicit baseline). Single-subject tests were conducted with data from subject 01 only, while group-level tests were run with all 14 subjects. Analyses were conducted in AFNI, SPM12, and FSL. In AFNI, single-subject variants were conducted using the uber_subject.py interface, which generates and runs two scripts: cmd.ap.sub_001 and proc.sub_001. Other variants did not require changing options in the interface, so were run directly from the command line, using a copy of the default cmd.ap.sub_001 script. Scripts for group-level tests were created manually. For each of the SPM variants, a batch.m file conducting the full analysis (using dependencies across processing steps) was created and run with the Batch Editor GUI. FSL-specific variants were modeled using FSL’s FMRI Expert Analysis Tool~( 9 ), where a.fsf file for the complete analysis was created using the FEAT GUI. Pre-defined Settings In this section, settings held constant over variants (e.g., drift modeling) are described for each of the packages. These pre-defined settings (including pre-processing) were identical for each variant. Pre-processing As slice-time information was not available for this study, this step was not considered in the