White matter microstructure alterations in frontotemporal dementia: Phenotype‐associated signatures and single‐subject interpretation

Abstract Background Frontotemporal dementias (FTD) include a genetically heterogeneous group of conditions with distinctive molecular, radiological and clinical features. The majority of radiology studies in FTD compare FTD subgroups to healthy controls to describe phenotype‐ or genotype‐associated imaging signatures. While the characterization of group‐specific imaging traits is academically important, the priority of clinical imaging is the meaningful interpretation of individual datasets. Methods To demonstrate the feasibility of single‐subject magnetic resonance imaging (MRI) interpretation, we have evaluated the white matter profile of 60 patients across the clinical spectrum of FTD. A z‐score‐based approach was implemented, where the diffusivity metrics of individual patients were appraised with reference to demographically matched healthy controls. Fifty white matter tracts were systematically evaluated in each subject with reference to normative data. Results The z‐score‐based approach successfully detected white matter pathology in single subjects, and group‐level inferences were analogous to the outputs of standard track‐based spatial statistics. Conclusions Our findings suggest that it is possible to meaningfully evaluate the diffusion profile of single FTD patients if large normative datasets are available. In contrast to the visual review of FLAIR and T2‐weighted images, computational imaging offers objective, quantitative insights into white matter integrity changes even at single‐subject level.


INTRODUCTION
White matter (WM) changes in frontotemporal dementia (FTD) have been extensively studied, and both clinical subtypes (Agosta et al., 2012;Borroni et al., 2007;Lam et al., 2014;Mahoney et al., 2014;Schwindt et al., 2013;Whitwell et al., 2010) and genotypes Rohrer et al., 2010) have been linked to relatively specific WM signatures. The most commonly utilized WM technique is diffusion tensor imaging (DTI), but a variety of non-Gaussian techniques such as diffusion kurtosis imaging (DKI) and neurite orientation dispersion and density imaging (NODDI) have also been successfully utilized (Wen et al., 2019). WM alterations in FTD can already be detected in the presymptomatic phase of the disease, and WM alterations are relatively marked by the time the diagnosis can be established (Wen et al., 2019). WM changes can also be readily tracked longitudinally across multiple timepoints to appraise the rate of progression and patterns of anatomical propagation. A shortcoming of descriptive imaging studies in FTD is that often only group-level inferences are presented, that is, shared patterns of WM disease burden in specific phenotypes or genotypes. The demands of clinical imaging differ significantly from the deliverables of academic radiology . The emphasis in the clinical setting is the accurate categorization of a suspected patient into a diagnostic subgroup, the evaluation of an asymptomatic mutation carrier with regard to presymptomatic disease burden, or the follow-up of a specific patient with an established diagnosis to verify if further pathology has been accrued Chipika, Siah, et al., 2020). The gap between group-level imaging and single-subject imaging is considerable in terms of practical utility, methodological challenges and academic relevance (Verstraete et al., 2015). While patterns of gray matter atrophy can be assessed in a variety of ways, the interpretation of single-subject WM profiles is particularly challenging. The visual inspection of FLAIR and T2-wighted images offers limited information, and the visual review of DTI data only permits the appreciation of movement, susceptibility or eddy-current related artefacts. In current clinical practice, the primary role of magnetic resonance imaging (MRI) is the exclusion of neoplastic, paraneoplastic, inflammatory and structural mimics rather than the confirmation of FTD-associated changes. Existing frameworks for single-subject categorization rely on various machine learning algorithms to classify single individuals into groups. A variety of supervised and unsupervised methods have been previously implemented across the spectrum of ALS-FTD. Models such as support vector machines, decision trees, neural networks and discriminant function analyses have been applied to imaging datasets with varying accuracy Bede, Murad, et al., 2021;Cajanus et al., 2018;Chagué et al., 2020;Davatzikos et al., 2008;Donnelly-Kehoe et al., 2019;Egger et al., 2020;Feis et al., 2018Feis et al., , 2020Grollemund et al., 2019;Klöppel et al., 2015;Koikkalainen et al., 2016;Manera et al., 2021;Schuster et al., 2016Schuster et al., , 2017Tong et al., 2017;Torso et al., 2020;Vernooij et al., 2018;Young et al., 2018). A common application of these approaches is the categorization of patients into FTD versus AD diagnostic groups. Bron et al., 2017;Canu et al., 2017;Frings et al., 2014;Hu et al., 2020;Kim et al., 2019;Ma et al., 2020;Möller et al., 2016;Wang et al., 2016;Yu et al., 2021). A key barrier to the development of successful machine learning algorithms in neurodegenerative conditions is the scarcity of uniformly acquired training data, especially in low-incidence phenotypes such as ALS-FTD, PLS-FTD, and postpolio syndrome (Aho-Ozhan et al., 2016;Finegan et al., 2019Finegan et al., , 2021Li Hi Shing, Lope, McKenna, et al., 2021;Lule et al., 2010;McKenna, Corcia, et al., 2021;Pioro et al., 2020;Trojsi et al., 2015Trojsi et al., , 2019. Accordingly, the objective of this study is to pilot an alternative quantitative WM rating framework for single-subject diffusion data interpretation based on tractwise z-scoring of diffusivity metrics with reference to demographically matched controls.  Kenna et al., 2013), and expansions longer than 30 hexanucleotide repeats were considered pathological.

Data acquisition
A spin-echo echo planar imaging (SE-EPI) pulse sequence with a 32direction Stejskal-Tanner diffusion encoding scheme was used to

Diffusion-weighted data processing
Diffusion-weighted (DW) data were preprocessed within MRtrix3, including noise removal and removal of Gibb's Ringing Artifacts. The topup-eddy algorithm was utilized for corrections for eddy-induced distortions and subject movements as implemented in FSL. Bias correction was performed with the ANTs1.9 N4 method. Diffusion tensors were fitted within MRtrix3, and maps of fractional anisotropy (FA) and radial diffusivity (RD) were generated. Anatomical images were preprocessed using FMRIB's FSL6.0's fsl-anat algorithm, including brainextraction and biasfield-corrections.

Tract segmentation
As the main objective of the study was the detection of WM microstructure integrity changes in individual patients, our analyses were restricted to regions of FA reductions and foci of increased RD as these diffusivity shifts indicate pathologic processes. Tract-wise probabilities of presumed pathology in individual subjects were estimated based on reference normative data. First, each patient's and control's DW data were segmented into 50 WM tracts using a neuralnetwork-based algorithm, TractSeg, which, as opposed to atlas-based approaches, does not assume a common anatomy between subjects and relies on individual WM fiber bundles anatomy. Peaks of the spherical harmonic function were extracted at each voxel to inform Tract-Seg, which were calculated from fitting voxelwise constrained spherical deconvolution (CSD). CSD is an alternative to the tensor model to perform tractography, which has been shown to outperform the tensor model in regions of crossing fibers, among others. Response functions were estimated using the dhollander method as implemented in MRtrix3 from which fiber orientation distribution function (fODF) could be calculated. Given that DW shells were acquired (b = 1000 and b = 0), a multishell approach could be implemented. Resulting fODFs were normalized according to (Raffelt et al., 2017); spherical harmonic peaks were retrieved from the normalized measures, which then served as input values into TractSeg.

z-Score-based tract integrity evaluation
The concept behind the z-scored-based strategy is the ascertainment of affected fiber bundles in individual patients. WM tracts were rated in individual patients with reference to age-or sex-matched HCs. Only tracts exhibiting significant FA reductions and increased RD were considered ''affected.'' First, subject-specific FA and RD maps were created for the segmented tracts by inputting each subjects' individual FA/RD map into TractSeg and averaging the estimated values across each tract. Normative data from HCs were z-scored, and patient data were normalized with respect to the relevant control group. Single patients' tract profiles were then contrasted to normative data using nonparametric statistics. First, the number of HCs exhibiting lower FA and higher RD than the observed value in the patient was determined for each patient and each tract. This value was then divided by the number of HCs (i.e., 50 both for males and females) to obtain p-values.
Given that two tests were run (decreased FA/increased RD), tracts with p < .025 were considered significantly different.
Finally, group-level statistics were also derived from the z-scorebased strategy to aid cross-validation against the standard approach.
We tested which tracts were preferentially affected across the entire patient group. To quantify this probability, probability distributions were first created reflecting the number of false positives across the patient group (i.e., p-values of < .025 provided a random event). This was modeled as a binomial process: where X is the random variable (a scalar), n is the number of correctly segmented tracts in the control distribution and p is the probability of assigning significance to a tract's p-value (in our case 0.025). This process was repeated 100,000 times to provide a dense probability distribution. p-Values were then derived for each tract by counting how many values in the null distribution exceeded the sum of significant observations across the patient group and dividing that count by the number of iterations. To match the threshold used in the validation arm of the study, the most affected tracts were identified using a relatively stringent alpha-threshold of p < .01.

Cross-validation by standard tract-based statistics
To validate the z-score-based approach, the group-level outputs were compared to those of an established analysis pipeline, FMRIB FSL's tract-based spatial statistics (TBSS). The voxelwise diffusivity profile of the five FTD groups was contrasted to controls. In accordance with FSL's TBSS recommendations, processing included outliner removal, nonlinear registration to the FMRIB58FA template and application of that transformation to align all subjects' FA/RD images to the MNI152 1 mm standard space. Voxelwise group-comparisons were computed using FSL's randomise algorithm, a nonparametric permutation testing scheme, with 2D-optimized threshold-free cluster enhancement (TFCE) to control for the family-wise error rate (FWER). To highlight the most pertinent WM changes, a stringent alpha-threshold of p FWER < .01 was applied.

Demographics
Two-sample t-tests were run between each male/female patient group versus the male/female control groups to confirm age matching. No statistical difference was found between any of the patient and control groups suggesting appropriate age matching. Relevant descriptive and inferential statistics are provided in Table 1.

z-Score-based subject-level inferences
The z-score-based strategy has successfully captured relevant WM pathology in individual subjects in each of the 60 FTD patients.

DISCUSSION
We have  In a research setting, imaging traits are typically derived from contrasting a group of patients with a specific clinical profile or a specific mutation to a group of demographically matched controls. In bvFTD, progressive WM changes have been described in the uncinate fasciculus, cingulum and corpus callosum; and to a lesser extent in the anterior thalamic radiation, fornix and superior and inferior longitudinal fasci-culus in both hemispheres (Agosta et al., 2012;Borroni et al., 2007;Lam et al., 2014;Mahoney et al., 2014;Whitwell et al., 2010). Studies in nvfPPA captured preferential left-sided changes in the anterior thalamic radiation, uncinate and superior longitudinal fasciculus (Agosta et al., 2012;Schwindt et al., 2013;Whitwell et al., 2010), which become more prominent in the right hemisphere over time . In svPPA, left-hemispheric uncinate, arcuate and inferior longitudinal fasciculus (Agosta et al., 2012;Borroni et al., 2007;Schwindt et al., 2013;Whitwell et al., 2010)  alterations. The assessment of cortical gray matter changes has been previously tested in a similar framework . It is conceivable that additional imaging measures, such as basal ganglia volumes normalized for total intracranial volume (TIV), alternative WM metrics, metabolite ratios and network coherence indices could be interpreted in a similar framework with reference to normative data (Abidi et al., 2020(Abidi et al., , 2021Dukic et al., 2019;Feron et al., 2018;Nasseroleslami et al., 2019;Proudfoot et al., 2018) as well as cord parameters in ALS-FTD cohorts (Bede et al., 2012;El Mendili et al., 2019;Lebouteux et al., 2014;Querin et al., 2019;Querin et al., 2018). Finally, it is plausible that statistical outputs from imaging modalities can be integrated into larger biomarker panels, which would include quantitative serum, cerebrospinal fluid, electroencephalographic, magnetoencephalographic, proteomic and neuropsychological indices (Blasco et al., 2018;Christidi et al., 2019;Devos et al., 2019;Dukic et al., 2019;Nasseroleslami et al., 2019;Proudfoot et al., 2018).
While the group-level outputs of the z-scored-based strategy and TBSS are anatomically concordant, their sensitivity in detecting WM changes is different. It is noteworthy that FA on TBSS does not capture WM degeneration in svPPA and bvFTD even at p < .01 using the appropriate covariates. Using the tract-wise approach, FA reductions are readily detected in the anterior corpus callosum in bvFTD and in the left inferior occipito-frontal and left superior longitudinal fascicles in svFTD (Table 3). At an individual level, the z-score-based approach readily detects the degeneration of relevant WM tracts in these two groups, which may be ''averaged out'' by less severe cases in the group comparisons ( Figure 1). TBSS generates voxelwise statistical maps projected on a WM skeleton which can be thresholded at a specific p-value, but it is typically reviewed visually, that is, anything below that threshold is highlighted as ''affected'' with a color spectrum map. In contrast, the text outputs from the z-score approach offer a list of ''affected tracts'' which can be ranked in order of ''severity'' based on associated p-values.
Both the tractwise analyses and TBSS suggest that RD is more sensitive to detect WM alterations in FTD. Based on RD profiles, affected tracts in bvFTD include corpus callosum, corticospinal tract and a number of subcoritco-cortical projections such as the superior thalamic radiation, thalamo-premotor and striato-premotor fibers. The involvement of the corticospinal tract in bvFTD is of interest as another shared feature between ALS and FTD. The involvement of bundles linking subcortical and cortical regions supports previous findings  and highlights the contribution of subcortical pathology to clinical manifestations (Christidi et al., 2018). WM degeneration in svPPA not only includes the corpus callosum, cingulum and arcuate degeneration but the left-hemisphere predominant involvement of long association fibers and projections from the thalamus and striatum ( Table 3). The nfvPPA cohort exhibits widespread degeneration of both commissural and long association fibers with slight left hemispheric predominance in addition to thalamic and striatal projections.
The C9orf72 negative ALS-FTD cohort not only exhibits widespread WM pathology in core ALS-associated regions such as the corticospinal tracts and corpus callosum, but in line with more recent studies, in the cerebellar peduncles, long association fibers, arcuate fasciculus, uncinate and cingulum McKenna, Chipika, et al., 2021; (Table 2). WM degeneration in ALS-FTD patients carrying the GGGGCC hexanucleotide expansion is comparable to the anatomical patterns observed in C9orf72-negative patients, but is more readily detected by FA reductions (Table 2). These observations highlight that contrary to previous suggestions, severe frontotemporal degeneration and subcortical involvement in ALS are not unique to the C9orf72 genotype.
In the absence of accompanying post mortem and CSF data, the participants of this study were merely categorized clinically. FTD phenotypes arise from different underlying proteinopathies (Chare et al., 2014;Josephs et al., 2011;Snowden et al., 2007); ALS-FTD is primarily linked to pTDP-43 (Geser et al., 2011), svPPA is nearly always associated with underlying TDP-43-C pathological aggregates (Bocchetta et al., 2020), nfvPPA is commonly associated with 4R tau (Spinelli et al., 2017) and molecular findings in bvFTD are thought to be heterogeneous (Perry et al., 2017). There are a number of study limitations we need to acknowledge, chief of which is the limited normative data at our disposal. Larger reference datasets stratified for narrow age brackets would permit more precise data interpretation. In this pilot study, we have only evaluated two diffusivity indices, but other diffusivity metrics, such as AD  or non-Gaussian diffusivity measures (Broad et al., 2019), could also be incorporated in zscore models. Finally, this is merely a cross-sectional study to test a quantitative, single-subject data interpretation framework. The natural expansion of this study would be tracking single subjects longitudinally to test whether our approach captures expanding WM pathology in single subjects over time. Notwithstanding these limitations, our findings indicate that our strategy offers valuable clinical insights in single-subjects and may be potentially developed into a viable clinical and pharmaceutical trial applications.

CONCLUSIONS
Frontotemporal dementia is associated with subtype-specific WM signatures, and regional WM degeneration is a key contributor to phenotype-defining clinical manifestations. The early diagnosis of FTD soon after symptom onset is challenging, and the current clinical role of imaging is limited to the exclusion of alternative structural, inflammatory or neoplastic pathologies. As demonstrated, carefully designed computational pipelines enable the interpretation of individual diffusion datasets and the ascertainment of anatomical patterns of WM degeneration in vivo. The development, optimization and validation of similar imaging frameworks that categorize individual patients based on raw MR data should be a key research priority. These initiatives signal a departure from describing group-level signatures, and herald a paradigm shift to precision, individualized, computational radiology.

ACKNOWLEDGMENTS
We thank all participating patients and each healthy control for supporting this research study. Without their contribution this study would not have been possible. We also express our gratitude to the caregivers and family members of participating patients for facilitating attendance at our neuroimaging centre. We also thank all patients who had expressed interest in this research study, but could not participate

CONFLICT OF INTEREST
The authors declare no conflict interest.

DATA AVAILABILITY STATEMENT
The data are not publicly available due to privacy or ethical restrictions.