Neurodegenerative disease is a common pathological hallmark underlying various types of dementia, including Alzheimer’s disease (AD) and frontotemporal dementia (FTD) (Rosen et al., 2002; Whitwell et al., 2007; Erkkinen et al., 2018). Brain atrophy is a key feature of these diseases, characterized by progressive neuronal loss, cortical thinning, and sulcal widening (Im et al., 2008). These structural changes are closely associated with cognitive and behavioral decline, making the quantification of brain atrophy crucial for early diagnosis, progression monitoring, and therapeutic evaluation (Brambati et al., 2009; Marino et al., 2019).

Frontotemporal dementia (FTD) comprises several clinical subtypes, including behavioral variant FTD (bvFTD), semantic variant primary progressive aphasia (svPPA), and non-fluent variant primary progressive aphasia (nfvPPA), each with distinct neuroanatomical atrophy patterns that differ markedly from those observed in AD (Risacher and Saykin, 2013; Ghosh and Lippa, 2015; Erkkinen et al., 2018; Yu et al., 2021; Eldaief et al., 2023; Taylor et al., 2025). While AD is typically associated with atrophy in the hippocampus and medial temporal regions (Ferreira et al., 2017), FTD more often presents with frontal and anterior temporal lobe degeneration, with variability in regional involvement depending on the clinical subtype (Lu et al., 2013; Planche et al., 2023). Identifying and characterizing these subtype-specific patterns of atrophy is critical for improving the differential diagnosis between AD and FTD.

Traditional methods for assessing brain atrophy, such as visual atrophy scales by clinicians, rely on subjective ratings of atrophy based on visual inspection of structural MRI scans (Scheltens et al., 1995; Kipps et al., 2007). While these methods remain prevalent in clinical practice, they are limited by their labor-intensive, time-consuming, susceptibility to inter-rater variability, and lack sensitivity to subtle anatomical changes (Duara et al., 2010; Cavallin et al., 2012; Choe et al., 2024). To overcome these limitations, advanced techniques, such as the measurement of cortical thickness (Fischl and Dale, 2000; Chu et al., 2024) and voxel-based morphometry (VBM) (Ashburner and Friston, 2000; Rosen et al., 2002) have been introduced, offering automated and more quantitative assessments of atrophy. However, these methods often struggle with segmentation inaccuracy due to the subtle intensity differences between gray-white matter boundary, leading to inconsistencies across imaging protocols and scanners (Leung et al., 2010; Pagnozzi et al., 2019), and they typically require complex preprocessing pipelines and spatial normalization steps that can impede their clinical utility due to high computational demands (Tustison et al., 2019; Bhagwat et al., 2021; Zhou et al., 2022).

To address these challenges, cerebrospinal fluid (CSF) volumetry offers a practical alternative: its strong intensity contrast with surrounding tissues enables robust segmentation across sites, and focusing on sulcal and ventricular CSF provides a reliable proxy for cortical atrophy (Kochunov et al., 2008; Im et al., 2008; De Vis et al., 2016). Building on this rationale, we employed a previously proposed deep learning–based segmentation model (Lim et al., 2025) to automatically quantify CSF volumes in 3D T1-weighted MRI scans. This approach serves as a proxy for sulcal widening and ventricular enlargement—key markers of neural tissue loss—while providing a practical and scalable solution aligned with routine clinical workflows. It facilitates intuitive interpretation, rapid clinical integration, and consistent application across diverse imaging settings.

In this study, regional CSF volume measures derived from the model were used to evaluate differences in brain atrophy patterns across three clinical subtypes of frontotemporal dementia (FTD), cognitively unimpaired (CU) individuals, and patients with Alzheimer’s disease (DAT). We hypothesized that each FTD subtype would exhibit distinct CSF volume-based regional atrophy patterns that could be statistically distinguished from those of DAT and CU individuals.

In this study, regional atrophy patterns among FTD subtypes were investigated using a deep learning-based segmentation model that quantifies CSF volume. The findings revealed significant differences in W-scores among diagnostic groups, demonstrating the algorithm’s effectiveness in capturing region-specific atrophy patterns. These results confirmed the initial hypothesis that each FTD subtype exhibits distinct spatial patterns of brain atrophy. Such precise characterization of atrophy may have clinical utility in improving differential diagnosis and disease monitoring in patients with FTD.

The results revealed clear and distinct atrophy patterns across FTD subtypes. Compared to DAT, bvFTD showed notably greater atrophy prominently in bilateral frontal regions (large effect sizes, Cohen’s d > 1.2), with a slightly stronger involvement observed in the right hemisphere. Medium-sized effects were also present in bilateral parietal and temporal lobes, anterior lateral ventricles, and hippocampal regions. Interestingly, parietal lobe involvement in bvFTD, although traditionally less emphasized, could reflect more advanced disease stages or clinical heterogeneity within this cohort. In contrast, the nfvPPA subtype demonstrated significant atrophy predominantly in the left frontal (Cohen’s d > 1.2) and left parietal regions, consistent with known left hemisphere-dominant language impairments. Similarly, svPPA showed pronounced left-dominant atrophy in temporal lobes and hippocampal regions, with the left hippocampal region exhibiting notably large effect sizes (Cohen’s d > 1.2). Such marked left-sided hippocampal involvement in svPPA aligns closely with previous findings, highlighting medial temporal lobe involvement beyond classic AD pathology. These observations are consistent with previously reported subtype-specific atrophy patterns (Risacher and Saykin, 2013; Ghosh and Lippa, 2015; Eldaief et al., 2023; Taylor et al., 2025), further reinforcing the validity and clinical relevance of the proposed approach (see Supplementary Table 2).

The 3D surface maps also provided compelling visual validation of these subtype-specific patterns, clearly demonstrating left-lateralized atrophy in nfvPPA and svPPA and widespread, anterior-predominant atrophy in bvFTD. These visualizations enhance the interpretability and clinical relevance of the observed structural differences, emphasizing the spatial specificity achievable by this method.

To further ensure the robustness of these findings, an additional propensity score matching (PSM) analysis was conducted using age, sex, and education as matching factors. After excluding cases with missing education data, 331 CU and 331 DAT participants were matched to 342 FTD participants with complete data (Supplementary Table 3). This adjustment reduced group size disparities and allowed for more balanced comparisons. Some regional differences that were previously significant (e.g., left temporal region in bvFTD, right frontal and left parietal regions in nfvPPA, and left frontal region in svPPA) were no longer observed after matching (Supplementary Figure 2) and the overall effect sizes were somewhat reduced (Supplementary Figure 3). However, the spatial patterns of atrophy remained largely consistent with the primary analysis.

These findings extend previous work conducted in Alzheimer’s disease, where stage-specific atrophy patterns—particularly in bilateral temporal lobes, lateral ventricles, and hippocampal regions—were demonstrated across a cognitive continuum from cognitively unimpaired individuals to patients with mild cognitive impairment and DAT. In the current study, quantitative CSF-based measures derived from the algorithm revealed significant hemispheric asymmetries—particularly within the left hemisphere—across FTD subtypes when compared with CU and DAT groups. Importantly, these findings were achieved using a quantification approach distinct from conventional methods (Scheltens et al., 1995; Ashburner and Friston, 2000; Fischl and Dale, 2000; Rosen et al., 2002; Kipps et al., 2007), yet consistently capturing relevant structural disease signatures. The ability to detect such detailed patterns highlights this method’s potential as a robust structural biomarker, particularly valuable for clinical evaluations in cases involving focal atrophy of the frontal and temporal lobes. This ensures reliable, scalable analysis across different clinical settings and patient groups. Given its reliance on standard clinical imaging modalities, this approach could be easily integrated into routine neuroimaging workflows, supporting clinicians in the early and differential diagnosis of dementia subtypes. With further validation, the algorithm may also offer potential benchmarks for future adaptations to more accessible imaging modalities, such as 2D MRI or even CT scans.

A major strength of this study is the practical capability of the deep learning-based method to quantify brain atrophy directly from routine MRI scans without complex preprocessing. However, several limitations should be acknowledged. The analysis relied solely on CSF volume, age, and sex, without incorporating additional factors (Sakka et al., 2011; Zhang et al., 2022) or direct measures of neurodegeneration such as cortical thickness or voxel-based morphometry. To complement the primary analyses, cortical thickness data in a subset of participants were also examined, which revealed broadly consistent and biologically plausible regional patterns with CSF-based measures (see Supplementary Figures 4, 5). Future work may examine the combined use of CSF- and GM-based features to improve both anatomical specificity and clinical interpretability. Additionally, the lack of clinical and cognitive data limits structure–function interpretations, and the cross-sectional design prevents assessment of longitudinal changes. Finally, the single-center design restricts generalizability, highlighting the need for external validation. Nonetheless, our findings demonstrate robust subtype-specific atrophy patterns, underscoring the method’s clinical relevance and value for dementia research.

In summary, this study demonstrated significant differences in regional atrophy patterns across FTD subtypes using a CSF-based deep learning algorithm. By focusing on clinically relevant RoIs, this simple and practical approach facilitates intuitive interpretation, rapid clinical integration, and consistent application across diverse settings. With further validation, this method holds significant potential as a robust structural biomarker for enhancing differential diagnosis and monitoring disease progression.