Computed tomography-based nnU-Net for region-specific brain structural changes across the alzheimer's continuum and frontotemporal dementia subtypes.

Quantifying structural brain changes is critical for diagnosing and monitoring neurodegenerative diseases. Although magnetic resonance imaging (MRI) is the silver standard, limited accessibility and cost hamper routine use. We developed a deep learning-based framework using the nnU-Net for brain segmentation using computed tomography (CT) to assess cerebrospinal fluid (CSF) volume changes, as an indirect marker of tissue loss, and evaluated its utility across Alzheimer's disease (AD) stages and frontotemporal dementia (FTD) subtypes. We included 2357 participants: cognitively unimpaired (CU, n = 595), mild cognitive impairment (MCI, n = 954), dementia of Alzheimer's type (DAT, n = 663), and FTD subtypes (FTD, n = 145, behavioral variant FTD (bvFTD, n = 66), nonfluent variant primary progressive aphasia (nfvPPA, n = 29), and semantic variant PPA (svPPA, n = 50). CT-based segmentation was trained and validated using 3D T1-weighted MRI as reference. We assessed (1) segmentation accuracy via Dice similarity coefficients (DSCs), (2) reliability and precision using correlation and Bland-Altman analyses, and (3) clinical utility by identifying stage- and region-specific changes in CSF volumes. Key regions, including anterior and posterior lateral ventricles, showed DSCs above 0.93 and correlations ranging from 0.822 to 0.996. CT-based measurements revealed increasing CSF volumes from CU to DAT and distinct patterns of CSF volume enlargement across FTD subtypes. This framework enables accurate, reliable assessment of CSF volume changes as an indirect marker of atrophy, and supports early detection and differential diagnosis.