Scanning was done in a 4-Tesla Siemens/Varian MR whole-body imager at the
laboratory for fMRI, Robarts Research Institute, University of Western Ontario,
London, Canada. In each experiment, a time series of 726 whole-brain EPI images
(interleaved Echo Planar pulse sequence) with an inter-scan interval of 3 seconds was
acquired over a period of 6 sessions (Fig. 1c). Each scan consisted of 17 horizontal
slices of 6 mm thickness [echo time (TE) 15 ms, 64x64 matrix size, 3.75 x 3.75 mm
in-plane resolution, field of view (FOV) 24 cm]. In addition, a high-resolution
anatomical image [Flash imaging sequence, TR 11 ms, TE 5.6 ms, 256x256 matrix
size, 0.93 x 0.93 mm in-plane resolution] consisting of 64 slices separated by 3.3 mm,
was collected for each subject.
Data Analysis
Behavioral analysis
While subjects performed experiments in the scanner, behavioral parameters
reflecting the number of sets completed (accuracy) and the time taken to complete a
set (response time) were measured for every block. Performance improvements within
an experiment and across the normal and rotated settings were assessed using
repeated-measures ANOVA. Based on the performance measures, we identified two
learning periods - an early period comprising the first two sessions in an experiment
where subjects were still slower and inaccurate and a late period comprising the last
two sessions of an experiment where the subjects approached their maximal levels of
performance.
Image analysis
The imaging data were analyzed using SPM99 (Wellcome Department of Cognitive
Neurology, London). The functional images were reoriented to set the origin near the
intersection of the coronal plane through AC and the AC-PC line and then motion
correction was performed with respect to the first functional image in each session.
Anatomical image for each subject was co-registered with the first functional image
and then normalized to the T1 template from the International Consortium for Brain
Mapping (ICBM) Project. The resulting parameters were used for normalizing all the
functional images (Friston et al., 1995a) into Talairach stereotaxic space (Talairach
and Tournoux, 1988). Spatial smoothing with a gaussian kernel of 8 mm FWHM was
applied to the normalized images. The preprocessed data were analyzed using the
general linear model framework (Friston et al., 1995b). For each subject, the
experimental settings were modeled using boxcar functions convolved with the
canonical hemodynamic response function in a session separable model. Data from all
the six sessions was included. We report results from the early (first two sessions) and
late stages (last two sessions) as the main focus of the experiment was to examine
representational changes in sequence learning. Further, subject-specific variations in
learning were taken into account by including the behavioral parameters (accuracy
and response time) as user specified regressors. The regressors were constructed by
giving a normalized score (range 0-100) reflecting improvements in learning for the
sequence blocks. A score of 0 was assigned to the follow blocks as there was no
learning involved.
Group analysis was performed using the random effects approach (Penny & Holmes,
2003) as implemented in the SPM99 software. Contrast images computed from the
subject-specific models were entered into paired t-tests that accounted for the two
repeated measures from the subjects. This model allowed for variance to be similar