Frontotemporal Dementia (FTD) is a neurodegenerative disorder characterized by extensive atrophy in the frontal and temporal lobes of the brain as well as high cerebrovascular burden. While anatomical Magnetic Resonance Imaging (MRI) is well established for quantifying brain atrophy in FTD, the variability in (pre-)processing methods limit the generalizability and comparability of findings. This study systematically compared the robustness and sensitivity of multiple widely used neuroimaging metrics, namely Deformation-Based Morphometry (DBM), Voxel-Based Morphometry (VBM), Cortical Thickness (CT), and segmentation-based grey matter Volumes, in detecting atrophy across FTD subtypes. We processed 732 T1-weighted MRI scans from 156 participants with FTD and 139 healthy controls from the Frontotemporal Lobar Degeneration Neuroimaging Initiative using our in-house pipeline PELICAN (Dadar et al., 2025) for volumetric measures and Free Surfer version 7 (Fischl, 2012) for CT and grey matter segmentations. Visual quality control using consistent quality control images at each step of the pipelines revealed significantly higher failure rates for CT (38.52%) and Free Surfer segmentations (23.63%) relative to PELICAN’s volumetric measures (2.04% DBM, 3.05% VBM). Failure rates differed between FTD subtypes and were related to pathological burden. Particularly for Free Surfer, errors occurred predominantly in regions with high prevalence of atrophy and White Matter Hyperintensities. In PELICAN, the addition of a FTD-specific template as an intermediate step during nonlinear registration decreased the failure rates in this step in the FTD population. We then applied linear regression models to assess each metric’s sensitivity in detecting cross-sectional differences between FTD groups controls as well as linear mixed-effects models to determine which method is most sensitive to longitudinal anatomical changes. While CT yielded effect sizes comparable to VBM and DBM when analyzing the same subset of successfully processed scans, VBM and DBM demonstrated enhanced power to detect effects due to lower failure rates and higher participant retention in the full sample. Overall, we demonstrate that image processing methodology and pipeline selection profoundly influences effect sizes and statistical power to detect meaningful between-group differences or longitudinal changes. Volumetric measures (DBM and VBM) yielded sufficiently robust pipeline outcomes to maintain adequate statistical power for capturing atrophy patterns after quality control procedures.

Frontotemporal dementia (FTD) is a group of clinical syndromes characterized by progressive impairments in language and/or abnormal changes in behavior (Bang et al., 2015; Olney et al., 2017). This umbrella term encompasses behavioral-variant FTD (bvFTD) and two forms of primary progressive aphasia (PPA): semantic-variant PPA (svPPA) and non-fluent variant PPA (nfvPPA) (Gorno-Tempini et al., 2011; Rascovsky et al., 2011). FTD is characterized by extensive neurodegeneration in the frontal and temporal lobes of the brain (Mackenzie et al., 2009) as well as neuropathological abnormalities in the form of hyperphosphorylated protein accumulations, typically composed of tau or TDP-43 (Mackenzie et al., 2010; Rademakers et al., 2012). Patients also exhibit increased levels of cerebrovascular pathologies, including leukoaraiosis which manifests as White Matter Hyperintensities (WMHs) in Fluid-Attenuated Inversion Recovery (FLAIR) or T2-weighted MRI scans (Dadar, Mahmoud, et al., 2022). While the value of anatomical MRI for both diagnosis and study of FTD has been well established (Meeter et al., 2017), there is considerable variability in the methods applied to (pre-)process and analyze MRI scans in the literature, limiting the generalizability and comparability of findings.

Several neuroimaging modalities and morphometric analysis techniques have been employed to quantify brain atrophy in FTD, including volume-based measures based on grey matter segmentation (Fischl et al., 2002), voxel-based morphometry (VBM) (Ashburner & Friston, 2000), and deformation-based morphometry (DBM) (Ashburner et al., 1998), and surface-based measures such as cortical thickness (CT) (Dale et al., 1999; Fischl et al., 1999).

Deformation-based morphometry (DBM)

The principle of DBM is to warp each individual scan to a common template through linear and non-linear deformation following pre-processing of the native scan, where local shape differences between the two images (i.e., the participant's T1-weighted (T1w) image and the template) are captured in the deformation fields. The local deformation obtained from the non-linear transformations can then be used as a measure of tissue expansion or contraction by estimating the determinant of the Jacobian for each transform (Ashburner et al., 1998). Local contractions can be interpreted as shrinkage of tissue (atrophy), and local expansions are often related to tissue growth, or ventricular or sulcal enlargement (Chung et al., 2001). DBM provides signal for cortical and subcortical grey matter as well as white matter. As DBM only relies on precise registration and typically does not involve smoothing, it can be sensitive enough to detect subtle systematic structural differences (Cardenas et al., 2007). On the other hand, the absence of spatial smoothing means that registration errors tend to have a stronger impact in DBM than in VBM, as smoothing helps mitigate small errors (Ashburner & Friston, n.d.).

Voxel-based morphometry (VBM)

VBM is another widely used neuroimaging technique for voxel-wise estimation of regional tissue volumes (Ashburner & Friston, 2000). Same as DBM, VBM can be applied to cortical and subcortical grey matter as well as white matter. The standard workflow for VBM involves spatial normalization, tissue segmentation, and spatial smoothing, meaning VBM constitutes a combination of DBM followed by tissue segmentation and smoothing. To create DBM maps, individual scans are spatially normalized to a common stereotactic space using linear and nonlinear registration to a standard template. This step ensures voxel-wise comparability and corrects for global head size and shape-related differences. Next, tissue segmentation categorizes tissue into grey matter, white matter, and cerebrospinal fluid based on intensity values and derives tissue probability maps. Tissue probability maps are then nonlinearly transformed to the standard template space and modulated by the DBM maps to identify local atrophy or expansion in the individual scans. Lastly, the resulting maps undergo spatial smoothing using a Gaussian kernel. This is necessary to account for residual small-scale interindividual anatomical differences after spatial normalization and to ensure a normal distribution of residuals, enhancing the validity of parametric statistical tests (Ashburner & Friston, 2000; Kurth et al., 2015; Whitwell, 2009). The primary limitation of VBM is its reliance on tissue segmentation, whereby any tissue misclassification can lead to inaccurate VBM estimations (Dadar, Potvin, et al., 2021). Due to the increased prevalence of lesions WMHs in individuals with neurodegenerative disorders, these errors might occur more frequently in these populations, introducing a systematic bias in derived metrics (Dadar, Potvin, et al., 2021). Furthermore, subtle signals might be removed during spatial smoothing, making VBM less sensitive to small volumetric changes (Ashburner & Friston, 2000; Chung et al., 2001). Lastly, despite it often being referred to as a metric of tissue density, the biological interpretation of VBM is unclear and should not be confused with cell density assessed cytoarchitectonically, as VBM is computed as the product of tissue classification probability and local volume change (Mechelli et al., 2005; Schwarz & Kašpárek, 2011). 

Free Surfer Cortical Thickness (CT) and Segmentatio

CT is defined as the distance between the inside and outside cortical surfaces, i.e. the boundary between white matter and grey matter and the grey matter/pial surface boundary (Lerch, 2015). The most common method of mapping CT uses surface-based techniques whereby preprocessing steps include linear registration to stereotactic space, nonuniformity correction, and tissue classification (Lerch, 2015). CT is then estimated by creating two polyhedral surfaces along the inside and outside cortical boundaries and calculating the distance between these two surfaces at each vertex. Postprocessing involves nonlinear surface-based alignment, parcellation of the cortex, and surface-based smoothing of the thickness maps along the surfaces (Dale et al., 1999; Fischl, 2012; Fischl et al., 1999). For longitudinal data, FreeSurfer first performs initial preprocessing on each time point and then builds a within-subject template that represents the subject’s average anatomy. All preprocessing steps are then repeated for each scan, using this template as a common reference, to improve robustness and reduce bias (Reuter et al., 2012).

This method allows for easy vertex-wise comparison of CT across populations, which might be more anatomically meaningful than comparing intensity values. The main downsides are the computational complexity (Scanlon et al., 2011) and, similar to VBM, its dependency on initial tissue classification, so that tissue misclassifications can lead to downstream errors. Additionally, CT does not provide measurements of subcortical or white matter atrophy. Given that evidence suggests subcortical atrophy often precedes cortical degeneration in FTD (Planche et al., 2023) and may serve as an important imaging biomarker of disease progression (Manera et al., 2022), the exclusion of subcortical measures represents a significant limitation.

In addition to CT metrics, FreeSurfer also explicitly segments and provides volumetric measures for cortical and subcortical grey matter structures. Following the aforementioned preprocessing steps, FreeSurfer combines voxel intensity information with spatial priors to assign each voxel a neuroanatomical label using a Bayesian classification framework. Spatial priors are derived from a probabilistic atlas constructed from manually labeled training data (Fischl et al., 2002). This is a significant limitation given that the probabilistic atlas was trained on relatively small, healthy samples, potentially reducing the accuracy in populations with atrophy, lesions (Dadar, Potvin, et al., 2021), or atypical anatomy (King et al., 2020). Consequently, FreeSurfer segmentations need to undergo careful and time-intensive quality control (Vahermaa et al., 2023).

Previous MRI studies on FTD have primarily focused on assessing brain atrophy through CT and VBM, demonstrating subtype-specific patterns of neurodegeneration (Chu et al., 2024; Meeter et al., 2017). BvFTD has been related to atrophy in frontal and insular cortices as well as the basal ganglia (Pan et al., 2012; Rosen et al., 2005; Schroeter et al., 2014; Seeley et al., 2008). SvPPA has been associated with left anterior temporal pole and hippocampal atrophy (Chan et al., 2001; Davies et al., 2009; Galton et al., 2001). In nfvPPA, atrophy seems to be most prevalent in the left inferior frontal gyrus, specifically involving Broca’s area (Akhmadullina et al., 2024; Bisenius et al., 2016; Mesulam et al., 2009; Rogalski et al., 2014). DBM has been applied in relatively few FTD studies, with results generally consistent with other MRI-based research (Cardenas et al., 2007; Manera et al., 2019; Shafiei et al., 2023; Wisse et al., 2021).

Given that VBM, DBM, and CT each have their own advantages and drawbacks, it remains unclear which of these three approaches would be best suited for studying FTD. While studies have compared the utility of different methods in other diseases, such as epilepsy (Scanlon et al., 2011) and multiple sclerosis (Righart et al., 2017), similar comparisons in the context of FTD are still lacking. Similarly, no systematic analysis of pipeline failure rates in FTD populations has been conducted. Considering that MRI processing workflows are typically optimized for healthy brains, they often struggle to accurately process scans with severe atrophy or lesions. This makes it essential to identify the method that can be most robustly applied to FTD individuals, who exhibit high levels of pathology, ensuring reliable and accurate imaging-derived measures in this population. These issues are especially important in clinical and longitudinal studies, where small participant samples lead to a lack of power to detect meaningful differences, and consistent detection of disease-related changes is essential for accurate diagnosis, patient stratification, and tracking disease progression.

To address these issues, we systematically compared the robustness and sensitivity of multiple widely used neuroimaging approaches (VBM, DBM, and FreeSurfer-derived CT and volume) in detecting atrophy across FTD subtypes. Using a cohort of individuals with clinically diagnosed FTD and cognitively unimpaired controls, we investigated which method results in the lowest pipeline failure rates while maintaining sufficient sensitivity to detect meaningful anatomical differences between FTD and healthy controls, as well as between FTD subtypes. Based on findings in other populations, we hypothesized that CT would be most sensitive in detecting distinctive patterns of cortical and subcortical atrophy in FTD and that DBM would achieve lower quality control failure rates than other techniques. Our goal was to provide an empirical foundation for selecting appropriate imaging tools in FTD research and clinical settings, and to highlight trade-offs between sensitivity and robustness that may influence methodological choices in future studies of neurodegenerative disease.

Participants

The Frontotemporal Lobar Degeneration Neuroimaging Initiative (FTLDNI, also referred to as NIFD) was founded through the National Institute of Aging and started in 2010 (https://memory.ucsf.edu/research/studies/nifd). The primary goals of NIFD are to identify neuroimaging modalities and methods of analysis for tracking frontotemporal lobar degeneration and to assess the value of imaging versus other biomarkers in diagnostic roles. NIFD provides MRI as well as various clinical and cognitive data. The inclusion criteria for FTD patients were a diagnosis of possible or probable FTD according to the FTD consortium criteria (Gorno-Tempini et al., 2011; Rascovsky et al., 2011). Longitudinal data from 292 NIFD participants were included in this project. Data was accessed and downloaded through the LONI platform in July 2023. The cohort includes patients with bvFTD (nbaseline = 77), svPPA (nbaseline = 39), and nfvPPA (nbaseline = 40) as well as 136 healthy controls (see Table 1 for demographic and cognitive characteristics of this cohort). In total, 766 T1w MRI scans were analyzed.

Table 1: Baseline demographic and cognitive characteristics in FTD subtypes and healthy controls in the NIFD cohort. Values expressed as mean (standard deviation). Asterisks indicate significant group differences based on one-way ANOVA or χ2 analysis comparing the groups. bvFTD: behavioral-variant Frontotemporal Dementia, svPPA: semantic-variant Primary Progressive Aphasia, nfvPPA: nonfluent-variant Primary Progressive Aphasia.

Acquisition of MR images

The FTLDNI uses the infrastructure established by the Alzheimer's Disease Neuroimaging Initiative (ADNI). All participating imaging centers share a common platform. Available information on acquisition parameters and scanners is summarized in Supplementary Table 1. For further details on MRI acquisition protocols and scanner information, please refer to https://cind.ucsf.edu/research/grants/frontotemporal-lobar-degeneration-neuroimaging-initiative-0.

MRI processing

PELICAN (VBM, DBM)

We utilized the Pipeline for Evaluating Longitudinal Images of Cerebral ANatomy (PELICAN) (Dadar et al., 2025), an open source extensively validated and widely used in-house pipeline for processing both cross-sectional and longitudinal imaging data (Fereshtehnejad et al., 2025; Kamal et al., 2025; Lajoie et al., 2025; Metz et al., 2024; Moqadam et al., 2024, 2025; Qiu et al., 2024). This pipeline is built on the open-source MNI MINC-Toolkit v2 (Neelin et al., 1998; Vincent et al., 2016) and ANTs (Avants et al., 2009) toolkits. Preprocessing steps include denoising with optimized non-local means filtering (Coupe et al., 2008), intensity inhomogeneity correction (Sled et al., 1998), and intensity normalization via linear histogram matching. Images are then linearly registered to the MNI152-2009c template (Fonov et al., 2011) (nine parameters: 3 translation, 3 rotation, and 3 scaling) using a hierarchical linear registration approach (Dadar, Fonov, et al., 2018). For longitudinal data, an individual-specific template is generated following these preprocessing steps. Subsequently, brain extraction from preprocessed T1w images is performed using the BEaST algorithm (Eskildsen et al., 2012). The extracted brains are then non-linearly registered to the MNI152-2009c template using ANTs nonlinear registration (Avants et al., 2009). DBM maps are derived by computing the determinant of the Jacobian deformation field at each voxel. Additionally, nonlinear registration is also performed with a specific FTD population template (Dadar, Manera, et al., 2021) as an intermediate registration target, and the final nonlinear transformation is calculated by concatenating the individual-to-indirect-template and indirect-template-to-average-template nonlinear transformations. This step was added to mitigate increased registration errors that can occur in populations with severe pathology when using average templates derived from healthy brains, such as the MNI152-2009c (Dadar, Fonov, et al., 2018; Dadar, Manera, et al., 2021; Dadar, Camicioli, et al., 2022). (Both direct and indirect registration pathways will be evaluated below.) Subsequently, the BISON pipeline (Dadar & Collins, 2021) is applied to the linearly registered images for tissue segmentation and WMH extraction. BISON enables robust tissue and WMH segmentation using either T1w images alone or a combination of T1w and either FLAIR or T2w images (Dadar, Maranzano, et al., 2018). Finally, VBM maps are created by multiplying the BISON tissue probability maps by the determinant of the Jacobian of the nonlinear deformation fields. To account for differences in head size between participants, similar to the approach used by Free Surfer (Klasson et al., 2018), we estimated Total Intracranial Volume (TIV) by dividing the total number of voxels in the MNI152-2009c template mask (i.e. the template TIV) by the scaling factor used to linearly align a participant’s scans to the standard template.

Scaling factor = x transform * y transform * z transform

Total Intracranial Volume (TIV) = total number of voxels in ICBM template mask (1,886,574 mm3)/scaling factor

Visual quality control was performed on multiple steps of the processing workflow, including linear and nonlinear registration, brain mask extraction, and tissue segmentation. VBM and DBM measures were calculated both voxel-wise and based on the CerebrA atlas, which includes 102 cortical, subcortical, and white matter regions (Manera et al., 2020). As the CerebrA atlas was derived from the Desikan-Killiany-Tourville (DKT) atlas (Desikan et al., 2006), which is used by FreeSurfer, regional results from PELICAN are comparable with FreeSurfer regional outputs.

Free Surfer (CT, Volume)

Estimation of CT and grey matter volumes was performed using longitudinal FreeSurfer (Fischl, 2012) version 7.4.1. The preprocessing pipeline includes individual-specific template creation, brain extraction, linear registration, and intensity normalization. Following preprocessing, the gray/white matter boundary is tessellated, and automated topology correction is applied. The surface is then deformed along intensity gradients to accurately position the gray/white and gray/CSF borders, generating the cortical models (Dale et al., 1999; Fischl et al., 1999). CT, i.e., the distance between the two cortical boundaries, is then computed at the vertex level. We quality-controlled the reconstructed gray/white and gray/CSF borders, separately for anterior and posterior regions and left and right hemispheres, using freeview visualizations. Here, the borders were overlaid on the preprocessed T1w images, and the rater consistently scrolled through the slices in axial (top to bottom), coronal (anterior to posterior), and sagittal (right to left) views. Mismatches between the estimated and the actual gray/white or gray/CSF borders were noted overall, and separately for left/right and anterior/posterior regions. CT measures were extracted both vertex-wise and parcellated according to the Desikan-Killiany-Tourville atlas (Desikan et al., 2006).

Additionally, we analyzed FreeSurfer grey matter segmentations. Following pre-processing, the pipeline performs tissue segmentation using voxel intensities and tissue probabilities (Fischl et al., 2002). We visually quality-controlled linear registrations and tissue segmentations. Visual quality control was performed, similar to PELICAN, on linear registrations and tissue segmentations. Volumes were extracted based on the Desikan-Killiany-Tourville atlas (Desikan et al., 2006). To account for differences in head size between participants, we also extracted the estimated TIV value provided by FreeSurfer for each participant.

Statistical Analyses

Robustness of processing pipelines

Pipeline outputs underwent visual quality control, and failure rates across the three methodologies were compared using chi-square tests. Demographic (age and sex) and clinical characteristics of participants whose scans passed versus failed quality control were compared using one-way ANOVA for continuous variables and chi-square tests for categorical variables. For clinical differences in the FTD cohorts, we assessed disease severity based on the sum of boxes score of the Clinical Dementia Rating. We also compared grey matter volumes and White Matter Hyperintensity burden, which was extracted from FLAIR images using PELICAN (BISON) (Dadar & Collins, 2021).

Correlation between methods

To assess the concordance between morphometric estimates derived from the different image-processing approaches, we conducted correlation analyses between regional measures obtained from PELICAN and FreeSurfer for the baseline visit of each participant. To achieve this, we computed region-wise partial correlations. For each brain region and pair of imaging methods, we modeled the association between the two regional measures both for the entire dataset as well as separately for cognitively normal participants and FTD patients. Within each group, we fit two linear regression models of the form Measure (method A) ~ Measure (method B) + Total Intracranial Volume (TIV)

For all combinations between DBM (indirect registration), DBM (direct registration), VBM, CT, and segmentation-based volumes. All measures and TIV were z-scored. Residuals from these models, representing variation in each measure unexplained by total intracranial volume, were extracted and correlated using Spearman correlation. This yielded a partial Spearman correlation for each region and method pair within the complete sample and each diagnostic group. Region-wise correlations were subsequently aggregated using Fisher's z-transformation to obtain group-level summary statistics. Furthermore, we evaluated the agreement between TIV estimates of PELICAN and FreeSurfer using Spearman correlation.

Sensitivity to group differences

To assess the sensitivity of each method in detecting differences in brain atrophy between FTD and healthy participants, as well as across FTD subtypes, we quantified and compared the effect sizes of diagnostic group differences across methods. In the case of DBM, analyses were conducted separately for measures derived using direct registration to the average template and those obtained via indirect registration using a disease-specific template. The following linear regression models were applied to cross-sectional data, separately for each neuroimaging method, comparing atlas-based and voxel-wise/vertex-wise atrophy measures across different regions: Measure ~1 + Diagnostic Group + Age + Sex + Total Intracranial Volume (TIV)

where the term Measure indicated the z-scored atrophy values derived from each method at baseline visits. Diagnostic Group was the variable of interest, distinguishing healthy controls from FTD patients. Age, Sex, and TIV (Brzezinski-Rittner et al., 2025) were added as covariates. Results were corrected for multiple comparisons using False Discovery Rate (FDR)(Genovese et al., 2002) correction with a significance threshold of p ≤ 0.05.

To ensure valid comparisons between methods while accounting for how increasing sample size with more robust pipelines affects statistical power, we conducted the analysis in two ways. First, we included all participants and scans that met quality control standards for each respective method, and second, we included only participants and scans that passed visual quality control across all methods. Numbers of participants/scans included in each part of the analyses can be found in Supplementary Figure 2.

Sensitivity to temporal change in brain atrophy

In order to assess the differences between the neuroimaging techniques in terms of detecting neurodegenerative changes in brain structure over time, the following linear mixed-effects models were applied, separately for each neuroimaging method, comparing atlas-based and voxel-wise atrophy measures across different regions:

Measure ∼ 1 + Diagnostic Group: Time from Baseline + Diagnostic Group + Time from Baseline + Diagnostic Group: Age at Baseline + Age at Baseline + Sex + Total Intracranial Volume (TIV) + (1|Participant ID)

where the term Measure indicated the z-scored values derived from each method at different participant timepoints. The interaction between Diagnostic Group and Time from Baseline (Diagnostic Group: Time from Baseline) was the term of interest, assessing the longitudinal slope in each FTD variant relative to healthy controls. Sex and TIV were added as covariates. Results were corrected for multiple comparisons using FDR (Benjamini & Hochberg, 1995; Genovese et al., 2002) correction with a significance threshold of p ≤ 0.05. Again, the analysis was performed both in all participants and in the subset of participants that met quality control standards across all methods. We also excluded follow-up visits that occurred later than 3.9 years after the baseline visit (90th percentile), to ensure a similar distribution of follow-up times between participants.

Quality control failure rates

Table 2: Results of quality control of MRI processing methods. Numbers and percentages of failed cases. VBM: Voxel-Based morphometry, DBM: Deformation-Based Morphometry, CT: Cortical Thickness, WMH: White Matter Hypointensity, BISON: Brain tissue segmentation pipeline (Dadar & Collins, 2021).

After visually assessing raw MRI scans and excluding scans with significant artifacts due to motion or incidental findings (e.g. stroke), a total of 732 scans remained for analysis. Run times varied considerably between the image processing methods. PELICAN processed participants with a single scan within 2 hours and required another 1.5-2 hours for each additional scan. FreeSurfer proved substantially more time-intensive, requiring up to 45 hours for a participant's initial two timepoints and an additional 15 hours for each subsequent scan.

Examples of quality control images, both successful and failed, are included in Figure 1 and Supplementary Figure S4. Quality control pass and fail rates for VBM, DBM, CT, and FreeSurfer segmentations are shown in Table 2. Using PELICAN, all scans passed brain mask creation. Only 3 scans failed linear registration to the MNI template. However, 50 scans (6.83%) failed direct nonlinear alignment to the average template, whereas only 15 (2.05%) failed nonlinear registration using the indirect FTD-specific average template. Notably, the majority of failed scans in direct nonlinear registration were from FTD participants (80.77% of failures), while failures in indirect nonlinear registration mostly occurred in healthy controls (80%) (Figure 3). Additionally, 12 (1.64%) scans resulted in erroneous tissue segmentations. Overall, DBM processing failed in 50 cases (6.83%) using direct nonlinear registration and 15 cases (2.04%) using indirect nonlinear registration, while VBM processing failed in 62 cases (6.56%) for direct and 25 cases (3.05%) for indirect nonlinear registration. Chi-square analyses revealed significant differences in failure rates between the use of direct and indirect registration to the MNI template both for VBM (p less than 0.001) and DBM (p less than 0.001) but no significant difference in robustness between DBM and VBM (p=0.122).

FreeSurfer failed to process 11 scans for segmentations and 5 for CT whereby the pipeline did not produce any outputs, despite multiple attempts to re-run. Among the scans that were fully processed, errors in linear registration occurred in 9 scans (1.25%) and tissue segmentation failed in 153 cases (21.22%). Quality control of CT maps showed misaligned grey matter/white matter or grey matter/pial surface estimations in 277 scans (38.10%) whereby 40-50% of patient scans were erroneous. Hence, FreeSurfer volume calculations failed for 173 scans (23.63%) and cortical thickness processing failed for 282 (38.52%) cases. Due to the high failure rate for FreeSurfer CT, we conducted regional quality control separately for anterior (frontal, temporal) and posterior (parietal, occipital) regions. The regional failure rates were: Anterior left: 157 scans (21.60%), anterior right: 159 scans (21.87%), posterior left: 129 scans (17.74%), posterior right: 141 scans (19.39%), whereby anterior regions showed more errors in patients, while posterior regions showed more errors in controls (Figure 3). Results of the chi-square analysis revealed that FreeSurfer volume processing yields lower failure rates than FreeSurfer CT (p less than 0.001). Furthermore, failure rates for DBM and VBM are significantly lower than either FreeSurfer volume or FreeSurfer CT (p less than 0.001).

Table 3: Demographic and clinical characteristics of participants whose scans failed in either one, both or neither of the tested neuroimaging pipelines FreeSurfer and PELICAN. Values expressed as mean (standard deviation). P-values are based on one-way ANOVA or χ2 analysis comparing the groups. Asterisks indicate significant group differences between the respective group and the “FreeSurfer & PELICAN Success” group based on t-tests.

When comparing demographic and clinical characteristics of participants whose scans passed versus failed visual quality control for either both or one of the pipelines, we found an association with disease severity as reflected by CDR sum of boxes, cerebrovascular burden, and sex (Table 3, Supplementary Figure S3), whereby scans from participants with higher Clinical Dementia Rating, higher WMH load, and male sex more frequently lead to processing errors.

Figure 1: Examples of passed (left) and failed (right) quality control images at different steps of the pipelines. Cyan arrows indicate areas of failure. PELICAN. Top row, left to right: 1. Raw T1w image. 2. Pre-processed T1w image after denoising, intensity inhomogeneity correction, and intensity normalization. 3. Contours of the MNI-ICBM152 average template overlaid in red on the linearly registered image, showing accurate alignment. 4. BEaST brain mask overlaid on the T1w image in green. 5. Contours of the MNI-ICBM152 average template overlaid in red on the non-linearly registered image using direct registration to the MNI-ICBM space, showing accurate alignment. 6. Contours of the MNI-ICBM152 average template overlaid in red on the non-linearly registered image using indirect registration to the MNI-ICBM space through the NIFD-FTD average, showing accurate alignment. 7. BISON tissue segmentation map overlaid on the preprocessed image. Bottom row, left to right, scans from different participants: 1. Raw T1w image, showing high amounts of motion that interfere with processing pipelines. 2. Pre-processed T1w image after denoising, intensity inhomogeneity correction, and intensity normalization, showing a failure in intensity normalization. 3. Contours of the MNI-ICBM152 average template overlaid in red on the linearly registered image, showing inaccurate alignment (skull included within the template space). 4. Contours of the MNI-ICBM152 average template overlaid in red on the non-linearly registered image using indirect registration to the MNI-ICBM space, showing inaccurate alignment of the ventricles. 5. Contours of the MNI-ICBM152 average template overlaid in red on the non-linearly registered image using indirect registration to the MNI-ICBM space through the NIFD-FTD average, showing inaccurate alignment in the occipital/temporal lobes. 6. & 7. Preprocessed T1w image and BISON tissue segmentation map overlaid on the same image, showing errors in tissue segmentation whereby White Matter Hypointensities were labelled as grey matter. FreeSurfer. Similar quality control images for different FreeSurfer steps. The Cortical Thickness panel shows FreeSurfer inner and outer surfaces overlaid on the T1w image in native space. Note the inaccuracies in the failed image (bottom row) indicated by cyan arrows.

Figure 2: Barplots summarizing the results of quality control of MRI processing methods per diagnostic group. Transparent bars indicate the total number of T1w MRI scans per group, saturated bars show the number of scans that failed visual quality control. Percentages denote the portion of scans per group per processing step that failed quality control. Upper row showing full results for each method, lower three rows showing quality control results for different steps of each pipeline. VBM: Voxel-Based Morphometry, DBM: Deformation-Based Morphometry, CT: Cortical Thickness, FS: FreeSurfer, bvFTD: behavioral-variant Frontotemporal Dementia, svPPA: semantic-variant Primary Progressive Aphasia, nfvPPA: nonfluent-variant Primary Progressive Aphasia.

Correlation between methods

We evaluated the inter-method agreement for regional metrics derived from DBM, VBM, CT, and segmentations based on the CerebrA/DKT atlas by calculating pair-wise partial correlations, adjusting for variation due to TIV (Figure 3A). The strongest correlations were observed among measures derived from PELICAN, whereby DBM using direct versus indirect registration as well as indirect DBM and VBM showed a mean correlation of r = 0.91, direct DBM versus VBM had a mean r = 0.83, and direct versus indirect VBM had a mean r = 0.87. In contrast, correlations between PELICAN- and FreeSurfer-derived volumetric measures (DBM/VBM versus FreeSurfer volumes) were lower, ranging from mean r = 0.46 to r = 0.51. Correlations between volume-based and surface-based metrics were generally low (mean r (Volume vs CT) = 0.37, mean r (VBM (direct) vs CT) = 0.35, mean r (VBM (indirect) vs CT) = 0.33). The lowest correlations overall were observed between DBM and CT (mean r = 0.21).

Figure 3B shows how the inter-method agreement differs among healthy controls and FTD patients. While we found that the patterns for PELICAN measure correspondence were overall similar between group, the mean correlation FreeSurfer outcomes were higher in the FTD group (e.g. Volume vs VBM (indirect): mean r (controls) = 0.32, mean r (FTD) = 0.55; CT vs VBM (indirect): mean r (controls) = 0.19, mean r (FTD) = 0.45).

As PELICAN and FreeSurfer both estimate TIV using the scaling factors during linear registration, we also compared their estimates (Figure 3C). While the correlation between both was high (r = 0.97), FreeSurfer estimates were consistently higher than PELICAN’s. This may result from differences in how FreeSurfer and PELICAN define intracranial space, or from the fact that FreeSurfer estimations are based on the MNI-305 template (https://surfer.nmr.mgh.harvard.edu/fswiki/eTIV), which is larger than MNI-ICBM152. Hence, FreeSurfer-based TIV estimates should not be compared to or combined with TIV estimates from other softwares.

Figure 3: Correlations between metric derived from different neuroimaging methods. A) Distribution of region-wise correlations between morphometric measures derived from DBM (direct or indirect nonlinear registration), VBM, Cortical Thickness, and FreeSurfer Volumes. Each datapoint represents a pair-wise Pearson’s correlation r-value for estimates of two methods for one of the CerebrA/DKT atlas regions. Mean r values were calculated using Fisher’s z-transform. B) Mean Spearman’s correlation r-value between regional measures of the different methods based on Fisher z-transform, divided by healthy controls (left) and FTD participants (right). C) Comparison between estimated Total Intracranial Volumes (in mm3) derived from FreeSurfer and from PELICAN. Each datapoint represents one participant and colours denote each individual’s sex. VBM: Voxel-Based Morphometry, DBM: Deformation-Based Morphometry, CT: Cortical Thickness, TIV: Total Intracranial Volume.

Sensitivity to group differences

The sensitivity of each method for detecting differences between FTD variants and healthy controls was assessed using linear regression models. In the regional analysis of the matched subset of scans successfully processed by all methods, VBM yielded the largest effect sizes (e.g., direct VBM: range(bv FTD) = [-0.97,-0.03], range(svPPA) = [-2.84,0.28], range(nfvPPA) = [-1.39,-0.02]) (Figure 4A/B, for t-statistics see Supplementary Figures S5/S6) across all FTD groups, followed by CT (range(bvFTD) = [-1.14,0.06], range(svPPA) = [-4.41,0.22], range(nfvPPA) = [-0.77,0.23]). Accordingly, VBM identified the most regions as significantly different after FDR correction (e.g. direct VBM n(bvFTD) = 62, n(svPPA) = 29, n(nfvPPA) = 61, Figure 4C), followed by DBM (e.g., direct DBM: n(bvFTD) = 32, n(svPPA) = 26, n(nfvPPA) = 28) and CT (n(bvFTD) = 42, n(svPPA) = 15, n(nfvPPA) = 17).

These patterns remained consistent in the full sample analysis (Figure 5A/B). However, the increased sample size and statistical power enabled VBM, DBM, and FreeSurfer Volumes to identify more significant regions, particularly in bvFTD and svPPA (Figure 5C), which had the highest failure rates and lowest numbers in the subset analysis. Sample size, and accordingly statistical power, for CT only marginally increased in the full sample analysis, so results for CT stayed widely consistent.

In the voxel-/vertex-wise analysis, VBM and CT identified similar atrophy patterns across FTD variants, with predominant frontal lobe atrophy in bvFTD, temporal lobe atrophy in svPPA, and scattered frontotemporal involvement in nfvPPA (Figure 6A). Furthermore, VBM and CT detected significant differences across a larger number of voxels/vertices than DBM (e.g., subset analysis in bvFTD: percent(direct DBM) = 6.6%, percent(indirect DBM) = 7.1%, percent(direct VBM) = 17.3%, percent(indirect VBM) = 29.1%, percent (CT) = 32.6%, Supplementary Figure 7), although DBM showed improved power and VBM outperformed CT in the full sample, whereby indirect nonlinear registration provided improved results (e.g. bvFTD: percent(direct DBM) = 25.3%, percent(indirect DBM) = 31.6%, percent(direct VBM) = 61.5%, percent(indirect VBM) = 69.2%, percent (CT) = 32.6%, Figure 6C). VBM

produced larger effect sizes than CT (e.g. subset analysis indirect VBM: range(bvFTD) = [-1.63,0.72], range(svPPA) = [-2.19,0.88], range(nfvPPA) = [-1.46,0.75]; subset analysis CT: range(bvFTD) = [-1.15, 1.08], range(svPPA) = [-2.53,0.72], range(nfvPPA) = [-1.07,1.27]), Figure 6B/Supplementary Figure 7, for t-statistics see Supplementary Figures S8/S9). DBM showed fewer significant voxelwise differences and notably identified effects in the opposite direction compared to other methods, particularly in periventricular regions adjacent to large sulci (see for instance hippocampus in svPPA).

As DBM and VBM also provide signal for volumetric changes in the white matter, we repeated the voxel-wise analysis in white matter areas for direct and indirect DBM/VBM (Figure 7). Patterns of white matter atrophy matched the locations of grey matter atrophy, whereby bvFTD showed bilateral frontal lobe atrophy, svPPA was focused on the temporal lobes, specifically the left hemisphere, and nfvPPA showed a more diffuse pattern including frontal and temporal lobes (Figure 7A). While indirect VBM generally identified more significant group differences, the performance differences between methods were less pronounced than in grey matter results, and beta estimate ranges were largely similar across methods (Figure 7B/C, for subset results see Supplementary Figure 10, for t-statistics see Supplementary Figure 11). 

Figure 4: Results of linear regression models assessing the sensitivity of each method to detect differences between FTD subtypes and healthy controls in the subset of MRI scans that were successfully processed by each method. A) Brain maps showing beta estimates for the main effect for diagnostic group, comparing regional values for FTD variants versus healthy controls. B) Box- and violin plots summarizing beta estimates for the group effect for each method. Each datapoint represents the estimate for one atlas region. C) Barplots showing the number of regions with significant group differences, after FDR correction, divided by direction of the effect. Percentages of significant regions out of all atlas regions, regardless of effect direction, are shown below each bar. *Note that Cortical Thickness is only calculated for 62 of 90 regions, as this measure does not include subcortical areas. VBM: Voxel-Based Morphometry, DBM: Deformation-Based Morphometry, by FTD: behavioral-variant Frontotemporal Dementia, svPPA: semantic-variant Primary Progressive Aphasia, nfvPPA: nonfluent-variant Primary Progressive Aphasia.

Figure 5: Results of linear regression models assessing the sensitivity of each method to detect differences between FTD subtypes and healthy controls in the full sample. A) Brain maps showing beta estimates for the main effect for diagnostic group, comparing regional values for FTD variants versus healthy controls. B) Box- and violin plots summarizing beta estimates for the group effect for each method. Each datapoint represents the estimate for one atlas region. C) Barplots showing the number of regions with significant group differences, after FDR correction, divided by direction of the effect. Percentages of significant regions out of all atlas regions, regardless of effect direction, are shown below each bar. *Note that Cortical Thickness is only calculated for 62 of 90 regions, as this measure does not include subcortical areas. VBM: Voxel-Based Morphometry, DBM: Deformation-Based Morphometry, bvFTD: behavioral-variant Frontotemporal Dementia, svPPA: semantic-variant Primary Progressive Aphasia, nfvPPA: nonfluent-variant Primary Progressive Aphasia.

Figure 6: Results of voxel-/vertex-wise linear regression models assessing the sensitivity of each method to detect differences between FTD subtypes and healthy controls in the full sample. A) Brain maps showing beta estimates for the main effect for diagnostic group in the grey matter, comparing voxel-/vertex-wise values for FTD variants versus healthy controls. B) Box- and violin plots summarizing beta estimates for the group effect for each method. Each datapoint represents the estimate for one voxel/vertex. For clearer visualization, 300 datapoints were sampled out of the distribution. C) Barplots showing the number of voxels or vertices with significant group differences, after FDR correction, divided by direction of the effect. Percentages of significant voxels or vertices, regardless of effect direction, are shown below each bar. *Note that Cortical Thickness is calculated for 327,684 vertices while VBM and DBM are extracted for 1,082,282 grey matter voxels. VBM: Voxel-Based Morphometry, DBM: Deformation-Based Morphometry, bvFTD: behavioral-variant Frontotemporal Dementia, svPPA: semantic-variant Primary Progressive Aphasia, nfvPPA: nonfluent-variant Primary Progressive Aphasia.

Figure 7: White matter results of voxel--wise linear regression models assessing the sensitivity of each method to detect differences between FTD subtypes and healthy controls in the full sample. A) Brain maps showing beta estimates for the main effect for diagnostic group in the grey matter, comparing voxel-wise values for FTD variants versus healthy controls.

B) Box- and violin plots summarizing beta estimates for the group effect for each method. Each datapoint represents the estimate for one voxel. For clearer visualization, 300 datapoints were sampled out of the distribution. C) Barplots showing the number of voxels with significant group differences, after FDR correction, divided by direction of the effect. Percentages of significant voxels, regardless of effect direction, are shown below each bar. VBM: Voxel-Based Morphometry, DBM: Deformation-Based Morphometry, bvFTD: behavioral-variant Frontotemporal Dementia, svPPA: semantic-variant Primary Progressive Aphasia, nfvPPA: nonfluent-variant Primary Progressive Aphasia.

Sensitivity to temporal change in brain atrophy

The sensitivity of each method for detecting longitudinal anatomical changes was assessed by comparing slope differences between FTD variants and healthy controls using linear mixed-effects models. The estimated slopes essentially indicate the increased rates of change for each FTD group that is above and beyond the estimated rate of change in the control group. In the regional analysis of the subset, VBM produced the largest effect sizes (e.g. direct VBM: range(bvFTD) = [-0.14,0.01], range(svPPA) = [-0.47,0.00], range(nfvPPA) = [-0.22,0.02], Figure 8A/B, for t-statistics see Supplementary Figures S12/S13) for most FTD groups, with the exception of svPPA, where CT detected the greatest slope differences in temporal cortex regions (e.g., beta(right entorhinal cortex, svPPA) = - 0.93, p less than  0.001). VBM again detected the highest number of significantly different regions after FDR correction (e.g. direct VBM n(bvFTD) = 56, n(svPPA) = 50, n(nfvPPA) = 48, Figure 8C), followed by DBM (e.g. direct DBM n(bvFTD) = 39, n(svPPA) = 49, n(nfvPPA) = 41) and FreeSurfer Volume (n(bvFTD) = 45, n(svPPA) = 47, n(nfvPPA) = 38). DBM exhibited some positive estimates, indicating less atrophy over time in patients versus controls, predominantly in periventricular regions such as the hippocampus (beta (left hippocampus, svPPA = 0.13, p less than 0.001). When analyzing the full sample (Figure 9A/B), these patterns persisted. The larger sample size and corresponding increase in statistical power allowed DBM, VBM an FreeSurfer Volume to detect a greater number of significant regions, with the most pronounced gains observed in bvFTD and svPPA (Figure 8C/9C).

Figure 8: Results of linear mixed-effects models assessing the sensitivity of each method to detect longitudinal changes between FTD subtypes and healthy controls in the subset of MRI scans that were successfully processed by each method. A) Brain maps showing beta estimates for the interaction between diagnostic group and time from baseline, comparing regional slopes for FTD variants versus healthy controls. B) Box- and violin plots summarizing beta estimates for the interaction effect for each method. Each datapoint represents the estimate for one atlas region. C) Barplots showing the number of regions with significant slope differences, after FDR correction, divided by direction of the effect. Percentages of significant regions out of all atlas regions, regardless of effect direction, are shown below each bar. *Note that Cortical Thickness is only calculated for 62 of 90 regions, as this measure does not include subcortical areas. VBM: Voxel-Based Morphometry, DBM: Deformation-Based Morphometry, bvFTD: behavioral-variant Frontotemporal Dementia, svPPA: semantic-variant Primary Progressive Aphasia, nfvPPA: nonfluent-variant Primary Progressive Aphasia.

Figure 9: Results of linear mixed-effects models assessing the sensitivity of each method to detect longitudinal changes between FTD subtypes and healthy controls in the full sample. A) Brain maps showing beta estimates for the interaction between diagnostic group and time from baseline, comparing regional slopes for FTD variants versus healthy controls. B) Box- and violin plots summarizing beta estimates for the interaction effect for each method. Each datapoint represents the estimate for one atlas region. C) Barplots showing the number of regions with significant slope differences, after FDR correction, divided by direction of the effect. Percentages of significant regions out of all atlas regions, regardless of effect direction, are shown below each bar. *Note that Cortical Thickness is only calculated for 62 of 90 regions, as this measure does not include subcortical areas. VBM: Voxel-Based Morphometry, DBM: Deformation-Based Morphometry, bvFTD: behavioral-variant Frontotemporal Dementia, svPPA: semantic-variant Primary Progressive Aphasia, nfvPPA: nonfluent-variant Primary Progressive Aphasia.

In this study, we conducted a systematic comparison between commonly applied neuroimaging methods and pipelines for estimating brain morphometry in an FTD cohort. First, pipeline robustness was assessed through rigorous visual quality control at each processing stage using consistent quality control images to identify processing failures. We found that failure rates increased with neurodegenerative and cerebrovascular lesion load in all methods, with notably higher failure rates observed in FreeSurfer. Errors occurred predominantly in areas with advanced atrophy or WMHs. Subsequently, we evaluated the sensitivity of each method for identifying anatomical differences between FTD subtypes and healthy controls, as well as detecting structural changes over time. While all methods revealed broadly consistent patterns of atrophy, VBM produced the highest effect sizes overall. Both DBM and VBM demonstrated superior statistical power for identifying subtle anatomical differences, attributable to lower processing failure rates and consequently larger sample sizes available for group-level analyses.

While PELICAN produced relatively robust outcomes (successful processing for direct DBM: 93.17%, indirect DBM: 97.95%, direct VBM: 91.80%, indirect VBM: 93.17%), FreeSurfer, despite being among the most popular choices, exhibited substantial failure rates (successful processing for Volumes: 77.53%, CT: 61.90%), leading to the exclusion of approximately one-third of scans from downstream analyses. This aligns with prior literature reporting high failure rates even in middle-aged populations (Monereo-Sánchez et al., 2021). Importantly, failure rates were significantly associated with disease severity and atrophy burden, meaning participants with the most pronounced and characteristic disease patterns had to be systematically excluded. These findings indicate that FreeSurfer has limited performance in populations expected to deviate substantially from the healthy templates that the software was trained on, including aging cohorts and dementia patients.

We observed that failure patterns predominantly corresponded to atrophy patterns, whereby higher failure rates occurred in frontal lobe regions for bvFTD and temporal lobe regions for svPPA. Similarly, Raamana et al. (2022) documented errors primarily in the temporal pole, with pial surface underestimation in 35% of cases and 80-100

FTLDNI MRI and clinical measures are available through https://ida.loni.usc.edu/login.jsp. PELICAN and FreeSurfer are available at https://github.com/VANDAlab/Preprocessing_Pipeline and https://surfer.nmr.mgh.harvard.edu/, respectively. Codes used for analysis are available on GitHub (https://github.com/ameliemetz/FTD_methods_comparison). Regional and voxel-/vertex-wise results are available for download under https://zenodo.org/records/17516941.

Amelie Metz: Conceptualization, Visualization, Writing - Original Draft, Formal analysis, Software

Roqaie Moqadam: Visualization, Software

Louis Collins: Conceptualization, Writing - Review & Editing

Yashar Zeighami: Conceptualization, Writing - Review & Editing

Sylvia Villeneuve: Conceptualization, Writing - Review & Editing

Mahsa Dadar: Conceptualization, Supervision, Writing - Review & Editing, Software

The authors report no competing interests.

The authors acknowledge Digital Research Alliance of Canada (https://www.alliancecan.ca/en) for the usage of the computing resources in the current work. Data collection and sharing for this project was also funded by the Frontotemporal Lobar Degeneration Neuroimaging Initiative (National Institutes of Health Grant R01 AG032306). The study is coordinated through the University of California, San Francisco, Memory and Aging Center. FTLDNI data are disseminated by the Laboratory for Neuro Imaging at the University of Southern California.