4.2.1 Data pre-processing

The data were individually corrected for head-motion using MCFLIRT. Mean-based intensity normalisation of all volumes by the same factor was applied (i.e. grand-mean scaling so that each of the 10 sessions had the same mean intensity value when averaged over 78 volumes and all brain voxels), followed by high-pass temporal filtering (see above) was performed. The individual data sets were registered into the space of the high resolution T1 image using FLIRT [Jenkinson and Smith, 2001]. In order to decrease computational load, the T1 high resolution image was segmented into different tissue types using FMRIB's Automated Segmentation Tool (FAST) [Zhang et al., 2001]. This provided maximum a-posteriori estimates for voxel-wise grey matter probability. Voxels with $p>0.2$ (N=47168; $\sim 18\%$ of all intra-cranial voxels) were included in the tensor analysis so that $\mbox{\protect\boldmath$X$}$ is a 3-way array of dimension $78\times 47168\times 10$. Based on the estimated sample covariance matrix of the $78\times 471680$ matrix $\mbox{\protect\boldmath$X$}_{I\times JK}$, the Laplace approximation to the model estimated a 19-dimensional signal sub-space. The data for each session was projected onto the space spanned by the first 19 Eigenvectors and spatially normalised by the voxel-wise variance estimate from the residuals of the projection.