Appendix e-1
Voxel Dimensions
Given the scarcity of patients available to perform studies of this type, and that scans of different patients were acquired at different times, it was not possible to fully control the parameters of the scan acquisition. As such, both FTD-U and FTD-T cohort scans present a mix of voxel dimensions, with voxel depth being either 1.5mm or 1.8mm (all other acquisition parameters were constant across patients). In order to maximise the number of scans available for the study, both voxel dimensions should ideally be used within the same analysis, while controlling for any nuisance effects that may follow. For this purpose, the recently describede-1 Interpolation Robustness (IR) metric was used. IR is a method of ensuring the statistical validity of combining data with differing voxel resolutions. IR was computed for each scan (Table e-1), using as reference the previously described N3+BET2. 4 random control patients and 1 FTD-U patient were down-sampled from 1.5mm to 1.8mm depth using a form of sinc interpolatione-2 using the mri_convert application available in the Freesurfer software package ( e-4. All were checked for post-interpolation cropping.
As suggested by a previous studye-1, balance of number of patients with 1.8mm voxel depth was enforced for contrasting cohorts by using an appropriate number of interpolated scans (mostly controls, except in analysis 4., in which 1 FTD-U degraded patient was used). The proportion of voxels with different dimensions can be seen in
Table e-1. A binary nuisance covariate modelling the difference in voxel dimensions was also used for all analyses, except in the PNFA, SD-T and FTD-A cases, given the limited number of patient scans (n = 3).
Population characteristics
The IR metric for contrasting groups was calculated (Table e-1) and tested for significance in order to validate the balance through interpolation approach. Differences were not significant for all relevant pairs of IR values, meaning that interpolation can be used in order to force balance without the risk of biasing the results
Additional Clinical Information
Two patients with FTD-U had a positive family history in a first degree relative: one (FTD-U with SD) had a sibling with motor neuron disease and the second (FTD-U with bvFTD) had a sibling with bvFTD and FTD-U pathology at post-mortem; neither had a progranulin gene mutation. All other patients had apparent sporadic disease. One patient with FTD-U/PNFA developed clinical motor neuron disease (MND) six months after scanning. Interestingly, four additional FTD-U patients developed dysphagia in the terminal phase of their illness though did not display overt clinical MND. Further information can be found in Table e-2.
Bias Correction and Skull Stripping
Recently, a studye-5 found that VBM using SPM5 can be improved by skull-stripping and bias-correcting MR images prior to being introduced in SPM5's generative model. In this study, all scans were pre-processed using the following automated pipeline: first, skull-stripping was performed using the hybrid watershed algorithm or HWAe-6 in FreeSurfer v.3.04 ( which integrates an atlas-based term constraining the shape of the brain; stripped volumes were then bias-corrected using the non-parametric non-uniform intensity normalisation or N3 v.1.10e-7 with default arguments; and finally, a fine brain extraction that excludes venous sinuses and cerebro-spinal fluid (CSF) was performed using the brain extraction tool v.2.1 or BET2e-8 in FSL v.3.3 ( with fractional intensity threshold, f, optimised for each patient individually using visual inspection of the results, and vertical gradient g set to zero.
Registration and Segmentation
After preprocessing scans were registered and segmented using the unified segmentation model provided in SPM5 ( Previous investigations have suggested that SPM5 is more accurate if it does not attempt to estimate bias fields when non-uniformities are not presente-9. Therefore, since r.f. bias was corrected by N3 during the pre-processing step, non-uniformity correction was disabled from the SPM5 model by providing parameter settings that caused a negligible effect over the volume of interest; that is, bias regularisation was set to 10 and bias FWHM was set to 150-mm cut-off. Default values were used for all other parameters. The default unified segmentation process was fully applied to the FTD-U scan in which the preprocessing step failed. Segments of grey matter (GM), white matter (WM) and cerebrospinal (CSF) were obtained in native space for all cohorts for the purpose of Total Intracranial Volume (TIV) estimation (see below). Unmodulated segments of GM, WM and CSF were obtained to build an explicit mask (see below). Modulated scans of GM were used for the statistical analyses. Modulation has the effect of preserving the total amount of GM represented by multiplying by the relative volumes. Finally, these segments were smoothed using an 8-mm full width at half maximum (FWHM) isotropic Gaussian kernel.
Total Intracranial Volume
A recent study demonstrated that TIV can be estimated reliably using SPM’s native space outputs of GM, WM and CSFe-10. These segments were therefore added and thresholded at 50% using Matlab7 (Mathworks Inc., Natick, MA, USA), and the result was multiplied by the individual voxel volumes in order to get the final estimate.
Non-Parametric Statistical Analysis
The PNFA and FTD-A cohorts had only three patients each and not very extensive abnormalities raising the possibility that the assumptions of normality of contrasting populations had been breached giving rise to false-positive results. Therefore, to confirm validity of the parametric results, the same analyses were performed using a permutation based, non-parametric approach as implemented in SnPM5 ( To compensate for the limited degrees of freedom, a 1000 permutation limit with 10mm FWHM variance smoothing was used as suggested in the SnPM manual. Grey matter segments were again thresholded at a relative level of 20%. Given the increased sensitivity of this method, a threshold of P (FWE corrected) = 0.05 was used. No extent threshold was applied.