Participants
A total of 271 individuals were included in the present study: 87 healthy subjects (HS), 31 semantic variant primary progressive aphasia (svPPA) [2], 37 behavioural-variant of fronto-temporal dementia (bvFTD) [1], 30 non-fluent/agrammatic variant primary progressive aphasia (nfvPPA) [2], 47 progressive supranuclear palsy Richardson's syndrome (PSP-RS) [3, 4] and 39 corticobasal syndrome (CBS) [5] cases. The FTD and control groups were from the frontotemporal lobar degeneration neuroimaging initiative (FTLDNI). PSP-RS and CBS groups were from the 4-Repeat Tau Neuroimaging Initiative (4RTNI, http://4rtni-ftldni.ini.usc.edu). The imaging and clinical methods are the same for both initiatives.
FTLDNI was founded through the National Institute of Aging and started in 2010. The primary aims of FTLDNI are to identify neuroimaging modalities and methods of analysis for tracking frontotemporal lobar degeneration (FTLD) and to compare the value of neuroimaging with other biomarkers in diagnostic roles. The Principal Investigator of FTLDNI is Dr. Howard Rosen (University of California, San Francisco). The data is the result of collaborative efforts at three different sites in North America. For more information, please visit http://memory.ucsf.edu/research/studies/nifd [https://ida.loni.usc.edu/collaboration/access/appLicense.jsp]. Access to the FTLDNI data was approved by the data access committee.
All subjects underwent an extensive clinical and neuropsychological evaluation and an MRI scan. The clinical diagnoses were made according to the current criteria [1,2,3,4,5]. Patients with vascular, psychiatric or other neurological disorders were excluded. All subjects included in the study were collected in the same centre with a similar acquisition protocol, in order to control potential confounding effects that could be caused by introducing variable acquisition protocols.
The sample was split, in order to have 2 different cohorts: the first group was used as a “Training cohort” to train a classification model (respectively 58 HS, 21 svPPA, 15 bvFTD, 17 nfvPPA, 24 CBS and 32 PSP-RS).
The second group was used as a “Test cohort” to test the model created in the training cohort, in an independent sample (respectively 29 HS, 16 svPPA, 16 bvFTD, 13 nfvPPA, 15 CBS and 15 PSP-RS).
Although the scans were acquired using the same acquisition protocol, a difference in the repetition time (TR 8200 or 6600) between subjects was found for a large portion of the cohort and this was used as a criterion to divide the sample into the 2 cohorts (Training and Test).
A subset of 8 individuals with autopsy confirmation (3 CBD and 5 PSP) from within the training cohort, was investigated separately to find the best cortical and subcortical features to differentiate the two groups. These features were used to create a model using the remaining subjects of the training cohort and then tested on the test cohort.
MRI data acquisition and pre-processing
MR images were acquired on a 3 Tesla Siemens Tim Trio system equipped with a 12-channel head coil at the UCSF Neuroscience Imaging Center, including the following acquisition:
Training cohort: (1) T1 MPRAGE (TR=2300 ms, TE=2.9 ms, matrix =240×256×160, slice thickness= 1 mm); (2) Diffusion sequences were acquired using the following parameters: TR/TE 8200/86 ms, b factor= 2000 s/mm2, isotropic voxels 2.2 mm3, 64 images with diffusion gradient applied in 64 non-collinear directions, 1 image with no diffusion weighting (b0).
Test cohort: (1) T1 MPRAGE (TR=2300 ms, TE=2.9 ms, matrix =240×256×160, slice thickness= 1 mm); (2) Diffusion sequences were acquired using the following parameters: TR/TE 6600/86 ms, b factor= 2000 s/mm2, isotropic voxels 2.2 mm3, 64 images with diffusion gradient applied in 64 non-collinear directions, 1 image with no diffusion weighting (b0).
The 3D T1-weighted image for each subject was segmented using the recon-all script included in Freesurfer v6.0 (http://surfer.nmr.mgh.harvard.edu/). The segmented masks were used to estimate the cortical thickness and the volumes of cortical grey matter (GM). The Brainstem Substructures tool [31] included in Freesurfer v6.0, was used for the segmentation of brainstem and four subregions (medulla oblongata, pons, superior cerebellar peduncle and midbrain). To account for head size, all volumes were normalised for total intracranial volume and expressed as fractions (fr).
All DTI images were processed using the FMRIB software library, (FSL Version 6.0.1, FMRIB, Oxford, UK, http://www.fmrib.ox.ac.uk/fsl/). Data was corrected for eddy and head motion and the diffusion tensor model at each voxel was fitted using DTIFIT. To control for the effect of head motion in DTI maps, a displacement index was calculated using an in-house script.
Cortical diffusivity analysis
Cortical diffusivity analysis was performed using novel software scripts. The software generates cortical profiles across the cortex in a radial direction, modelling the columnar organisation within the cortex (Fig. 1) [20, 21]. Values for the diffusion tensor derived metrics were averaged along the cortical profiles, across the entire grey matter mask. The metrics calculated were MD, and three measures relating to the principal diffusion component, namely the angle between the radial minicolumn direction across the cortical layers and the principal diffusion direction (AngleR, θrad); the principal diffusion component projected onto the plane perpendicular to the radial minicolumn direction across the cortex (PerpPD, (×10–3 mm2/s) and the principal diffusion component projected onto the radial minicolumn direction across the cortex (ParlPD, (×10–3 mm2/s)) [20, 21].
This set of cortical diffusivity measures have been previously correlated with amyloid and tau PET [20]. The correspondence between tau tangle pathology and cortical diffusivity has been demonstrated on a small scale by a correlation in post-mortem MRI scan data between cortical disarray measurement in the medial temporal lobe and Braak staging in histopathological microscope sections from the same subjects [32]. All the cortical values were averaged to reduce the influence of noise in the DTI scans, effectively smoothing the data, and ensuring only directionality with some local coherence would dominate, guarding against the influence of random deflections from the radial direction. Previous work has found that measures of the cyto- and myelo-architecture are relatively stable within a cortical subregion [33] indicating that it is valid to find an average value for that region.
The whole-brain DTI maps were used to extract a single value for each cortical region segmented using the recon-all pipeline of the FreeSurfer v6.0 software package (http://surfer.nmr.mgh.harvard.edu/) based on the Desikan-Killiany cortical atlas.
Design and statistical analysis
Data were analysed using IBM SPSS Statistics version 26 (SPSS, Chicago, IL). Normality was tested using Shapiro-Wilk tests.
Analysis of variance was performed by using the multivariate General Linear Model of SPSS to compare the diagnostic group differences in cortical diffusion measurements in our cohorts, using the diagnostic group code as a fixed factor and age, sex and head motion as covariates.
All statistically significant results reported remained significant after false discovery rate correction (FDR < 0.05) [34].
Feature reduction and classification
To assess the diagnostic group discrimination capability of the cortical diffusivity measures, in a binary (HS vs Patients) and multiclass (HS, bvFTD, svPPA, nfvPPA, PSP-RS and CBS) problem, linear discriminant analysis was used.
In the binary classification, the “best discriminator” to distinguish HS and patients was considered to be the measure with the highest accuracy statistic. Once identified in this manner, regional values of this measure were then used for the multiclass classification.
In the multiclass classification, a large number of initial features were reduced to improve the classification performance, removing irrelevant or redundant variables using principal components analysis (PCA) (SPSS Factor analysis) as a filter method. The components with a combined variance of 95% were selected for the multiclass classification.
Finally, to summarise the predictive value of each measurement, we also computed sensitivity (SENS), specificity (SPEC), positive and negative predictive values (PPV, NPV) and the positive and negative likelihood ratios (L+ and L−).
Autopsy confirmed sample
An autopsy confirmed diagnosis was available for 17 participants, and of those, 8 cases with confirmed CBD and PSP (3 CBD and 5 PSP) were included in analyses. These subjects were included in an additional investigation to assess the diagnostic accuracy of classifying CBS and PSP-RS using regional values of the best whole brain diffusion measure and subcortical volumes. This autopsy-confirmed sub-group was used to select a set of features that were used to train a linear discriminant analysis model in the training cohort,which was then tested in the test cohort. A PCA was performed to reduce the number of variables used in the linear discriminant analysis.