next up previous
Next: Inter-Session Evaluation Methods Up: Methods Previous: Original Experiments and Analysis

Methods Tested

We now describe the analysis approaches used for this paper. The two packages used for our investigations were SPM99b (Statistical Parametric Mapping, www.fil.ion.ucl.ac.uk/spm) and FSL (FMRIB's Software Library, www.fmrib.ox.ac.uk/fsl, version 1.3, June 2001). Both are freely available and widely used.

SPM includes a motion-correction (``realignment'') tool, a tool for registration (``normalisation'') to standard-space, GLM-based time-series statistics [23] and GRF-based inference [11]. SPM carries out standard-space registration before time-series statistics. The SPM99b time series statistics correct for temporal smoothness by precolouring [10].

GLM-based analysis in FSL is carried out with FEAT (FMRI Expert Analysis Tool), which uses other FSL tools such as BET (Brain Extraction Tool [20]), an affine registration tool (FLIRT - FMRIB's Linear Image Registration Tool [14,13]), and a motion-correction tool based on FLIRT (MCFLIRT [13]). FEAT carries out standard-space registration after time-series statistics. FSL time series statistics correct for temporal smoothness by applying prewhitening, as described in [22].

6 different, complete analyses were carried out with various combinations of pre-processing and time-series statistics options, in order to allow a variety of comparisons to be made. In tests A,C,G, FSL was used for preprocessing and registration whilst in tests D,E,F, SPM was used. For tests A,D,G, FEAT time-series statistics was used whilst for C,D,F, SPM time-series statistics was used.

In tests A-E the various controlling parameters were kept as similar as possible, both to each other and to default settings in the relevant software packages. Tests A vs D and C vs E hold the statistics method constant whilst comparing spatial methods, therefore showing the relative merits of the ``spatial'' components (motion correction and registration). Tests A vs C and D vs E hold the spatial method constant whilst comparing statistical components, thus showing the relative merits of the statistical components (time-series analysis). A vs E tests pure-FSL against pure-SPM. F and G test pure-SPM and pure-FSL respectively, with these analyses set up to match the specifications of the original analyses in [18] as closely as possible, including turning on intensity normalisation in both cases. For a summary, see Table 1.

(For B, ICA-based temporal-model-free analysis was carried out; the model-free results are not included in this paper, but will be presented elsewhere.)

Because the methods for high-pass temporal filtering in FSL and SPM are intrinsically different, they cannot be set to act in exactly the same way (within A-E and within F,G) by choosing the same cutoff period in each; instead, the cutoff choices were made so as to match as closely as possible the extent to which the relevant signal and noise frequencies were attenuated by the different methods. For the purposes of this paper, high-pass temporal filtering is considered to be part of the temporal statistics, where it most naturally fits.

The non-default ``Adjust for sampling errors'' motion-correction option in SPM was not used.

8 sessions (of the 99) were excluded from the original analysis in [18], due to ``obvious movement artefacts''. These were however included in our analyses as we did not consider that there was sufficient objective reason to exclude them; the estimated motions for these sessions were not in general significant outliers relative to the average motion across sessions and any apparent (activation map) motion artefacts were not in general significantly different from the majority of the sessions. The quantitative results given in Sections 4.3 and 4.6 were in fact recalculated without these 8 sessions (i.e., reproducing the same dataset as used in [18]), but without any significant change in results, and therefore are not reported here.


Table 1: Different analyses carried out.

Preprocessing Statistics Registration
A FSL FSL (FEAT) FSL (FLIRT)
(MCFLIRT spat=5 intnorm=n) (hp-FSL=40)
C FSL SPM FSL (FLIRT)
(MCFLIRT spat=5 intnorm=n) (hp-cos=72)
D SPM FSL (FEAT) SPM
(SPM-mc&norm spat=5 intnorm=n) (hp-FSL=40) (done in preproc)
E SPM SPM SPM
(SPM-mc&norm spat=5 intnorm=n) (hp-cos=72) (done in preproc)
F SPM match [18] SPM SPM
(SPM-mc&norm spat=6 intnorm=y) (hp-cos=98.4) (done in preproc)
G FSL match [18] FSL (FEAT) FSL (FLIRT)
(MCFLIRT spat=6 intnorm=y) (hp-FSL=53)


spat - spatial filtering with full-width-half-maximum given in mm.
intnorm - intensity normalisation - the intensity rescaling of each volume in a 4D FMRI data set so that all have the same mean within-brain intensity.
hp-FSL - FSL's high-pass temporal filtering with cutoff period given in seconds.
hp-cos - high-pass temporal filtering (in seconds) via cosine basis functions.



next up previous
Next: Inter-Session Evaluation Methods Up: Methods Previous: Original Experiments and Analysis