Demographics
MRI data obtained at 1.5 T during routine clinical neuroimaging were approved by the Cleveland Clinic Institutional Review Board for storage and analysis as de-identified images after patients provided verbal consent. The same data set used in our previous studies [6] was analyzed in 15 unaffected neurologic controls (10 men, 5 women) aged 57.1 ± 19.2 years (mean ± SD, range 28–95 years) identified at the time of MRI, and 18 patients with ALS-FTD (5 men, 13 women) aged 66.9 ± 10 years (mean ± SD, range 52–87 years) identified by bedside (Montreal Cognitive Assessment, MoCA) or formal neuropsychometric testing. Clinical details of the ALS-FTD patients, as defined by Neary criteria [9], and neurologic controls are provided in our previous report [10].
Image acquisition
T1-weighted data were obtained on a 1.5 T magnet (Siemens Symphony, Erlangen, Germany) using 3D magnetization-prepared rapid gradient echo (M-PRAGE) sequence. Imaging parameters were: 160 slices, voxel resolution of 1.0 × 1.0 × 1.0 mm; pulse sequence parameters were: TR = 1970 ms, TE = 4.38 ms, number of averages = 1, and scan time = 6.45 minutes.
Data processing
VBM analysis was carried out using FSL and SPM softwares separately as described below.
FSL approach
FSL’s standard VBM processing pipeline was adopted and the processing steps are briefly described below. An optimized VBM approach of Good et al. [10] was adopted with all processing steps carried out using openware FSL version 4.1.5 (http://www.fmrib.ox.ac.uk/fsl/) [9]. Data processing was divided into four major steps: 1) T1-weighted images were brain-extracted using BET [11] adopting the suggestions for using FSL’s BET outlined by Popescu et al. [12]. Any leftover non-brain regions of ALS-FTD patients were manually edited by painting the non-brain voxels with a mask of this created to exclude these non-brain voxels from the brain image. An experienced neurologist (EPP) with extensive neuroanatomical knowledge confirmed that only non-brain regions were removed by manual edits. Because ALS does not typically result in T1 hypointense lesions in the brain (as was the case in our ALS-FTD patients as well), no correction was applied to T1-weighted images before the segmentation step. 2) Brain extracted images were segmented into white matter, GM, and cerebrospinal fluid (CSF) volume probability maps using FAST [13]. 3) In order to avoid bias during the registration process, a study-specific GM template was created by registering into MNI152 space with the affine registration tool FLIRT [14,15]. A randomly chosen subset of subjects, as suggested in the FSL VBM user guide from both unaffected neurologic controls (n = 15) and ALS-FTD patients (15 subjects randomly chosen out of 18), was chosen to create the above study-specific template. After nonlinear registration using FNIRT (www.fmrib.ox.ac.uk/analysis/techrep), the resulting images were averaged to create the template. 4) All the native GM images were non-linearly reregistered to the template and modulated (affine component not included) using the Jacobian of the warp field. 5) These images were then smoothed using a full-width half-maximum (FWHM) of 7 mm, and 6) general linear model (GLM) was used to compare voxel-wise differences in GM volume between ALS-FTD and the control groups. Non-parametric statistics were performed using “randomise” with 5000 permutations and using threshold free cluster enhancement (TFCE) option either enabled or disabled (i.e. voxel-based thresholding without the TFCE option in randomise). Statistical parametric maps generated both with and without the TFCE option were then compared with SPM. Variance smoothing was not used.
SPM approach
VBM analysis in SPM8 software was carried out using VBM8 toolbox by adopting standard VBM processing routine. The processing steps are briefly explained below: 1) estimate and write, 2) DARTEL create template, 3) DARTEL existing template, 4) normalize to MNI space, and 5) non-parametric statistics. More specifically, the first step (estimate and write) involves bias-correcting the raw T1-weighted images for inhomogeneities, extracting the brain, and segmenting it into GM, WM and CSF volume probability maps. The DARTEL create template step was used to create a customized template for our study; the same randomly chosen subjects that were used in the FSL approach were used here too. Once the study-specific template was created from the above step, the remaining subjects were registered nonlinearly to this template using DARTEL existing template module [16]. After normalizing and registering all subjects to MNI space, the resulting images were modulated (without including affine component) and smoothed using a full-width half-maximum (FWHM) of 7 mm. Finally, the smoothed images were used for statistical inference. Statistical non-parametric mapping (SnPM) with 5000 permutations without variance smoothing was used to compare voxel-wise differences in GM volumes between the ALS-FTD and control groups.
Freesurfer approach
Cortical volume and thickness measures were estimated using the openware, Freesurfer (http://surfer.nmr.mgh.harvard.edu/). Our MR data were of high quality, and appropriate checks and edits were performed throughout the entire Freesurfer workflow (both authors evaluated/checked all the steps especially, brain segmentation, registration to Talaraich space, pial surface and GM-WM boundaries extraction results in each subject). Standard image-processing steps were adopted including: 1) correct for motion artifacts and strip skull based on a hybrid watershed/surface deformation procedure [17], 2) register images to a Talairach brain template and segment for subcortical WM and GM structures [18], 3) estimate the GM-WM boundary via a tessellation step, and subsequently perform automated topology correction, 4) optimally place GM-WM and GM-cerebrospinal fluid (GM-CSF) boundaries using surface normalization and intensity gradients, 6) after cortical models are complete, perform deformable procedures for further processing and analysis, such as surface inflation and registration to a spherical atlas [19]. Both intensity and continuity information from the entire 3D MR volume are used to produce representations of cortical thickness, where thickness is measured as the closest distance from GM-WM to GM-CSF boundary at each vertex on the tessellated surface [20].
Level of significance in Freesurfer, SPM and FSL was considered a p value <0.05 corrected for multiple comparisons using family-wise error rate. Covariate age was regressed out in the GLM. To quantitatively compare between SPM, FSL VBM results with Freesurfer’s cortical thickness and volume measures, we calculated percentage of voxels that reached statistical significance [6]. Similarity between statistical parametric maps of SPM, FSL with Freesurfer was obtained for motor and extra-motor regions using Dice similarity index [21-23]. In order to measure Dice similarity index, we voxelized cortical thickness and cortical volume surface statistical parametric maps to MNI space because FSL and SPM statistical parametric maps were already in MNI space. Comparing volumetric (VBM) and thickness (Freesurfer) parameters requires conversion between the two units of measure. We performed surface to volume map conversion, after Klein et al. [22], as only one resampling was necessary (i.e., Freesurfer statistical parametric maps) in comparison to volume to surface map conversion for which two resamplings would be needed (i.e., FSL and SPM statistical parametric maps). Statistical parametric maps of cortical thickness and cortical volume were sampled to the target volume (FSL and SPM statistical parametric maps were in MNI space) using mri_surf2ol, mri_aparac2aseg and mri_convert commands in Freesurfer. The resulting volume maps from Freesurfer were then used to measure Dice similarity indices with FSL and SPM statistical parametric maps. Dice similarity index measures similarity by taking the mathematical intersection (voxels common to both images in the given ROI) of similarly labeled regions (here motor cortex and non motor cortex ROI’s) between the two softwares and then dividing by the mean volumes of the two ROIs [22].