Improved Multiscale Structural Mapping with Supervertex Vision Transformer for the Detection of Alzheimer's Disease Neurodegeneration.

Improved Multiscale Structural Mapping with Supervertex Vision Transformer for the Detection of Alzheimer's Disease Neurodegeneration - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Search Log in Primary site navigation Search Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Hum Brain Mapp . 2026 Jun 2;47(8):e70548. doi: 10.1002/hbm.70548 Improved Multiscale Structural Mapping with Supervertex Vision Transformer for the Detection of Alzheimer's Disease Neurodegeneration Geonwoo Baek 1 , David H Salat 2, 3, 4 , Ikbeom Jang 1, ✉ ; for the Alzheimer's Disease Neuroimaging Initiative Author information Article notes Copyright and License information 1 Department of Computer Science and Engineering, Hankuk University of Foreign Studies, Seoul, Republic of Korea 2 Athinoula A. Martinos Center for Biomedical Imaging, Department of Radiology, Massachusetts General Hospital, Charlestown, Massachusetts, USA 3 Department of Radiology, Harvard Medical School, Boston, Massachusetts, USA 4 Neuroimaging Research for Veterans (NeRVe) Center, VA Boston Healthcare System, Boston, Massachusetts, USA * Correspondence: Ikbeom Jang ( ijang@hufs.ac.kr ) ✉ Corresponding author. Revised 2026 Apr 6; Received 2025 Dec 12; Accepted 2026 May 5; Collection date 2026 Jun 1. © 2026 The Author(s). Human Brain Mapping published by Wiley Periodicals LLC. This is an open access article under the terms of the http://creativecommons.org/licenses/by-nc/4.0/ License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited and is not used for commercial purposes. PMC Copyright notice PMCID: PMC13240337  PMID: 42227644 ABSTRACT Alzheimer's disease (AD) confirmation often relies on positron emission tomography (PET) or cerebrospinal fluid (CSF) analysis, which are costly and invasive. Consequently, structural MRI biomarkers such as cortical thickness (CT) are widely used for noninvasive AD screening. Multiscale structural mapping (MSSM) was recently proposed to integrate gray–white matter contrasts (GWCs) with CT from a single T1‐weighted MRI (T1w) scan. Building on this framework, we propose MSSM+, together with surface supervertex mapping (SSVM) and a Supervertex Vision Transformer (SV‐ViT). 3D T1w images from individuals with AD and cognitively normal (CN) controls were analyzed. MSSM+ extends MSSM by incorporating sulcal depth and cortical curvature at the vertex level. SSVM partitions the cortical surface into supervertices (surface patches) that effectively represent inter‐ and intra‐regional spatial relationships. SV‐ViT is a Vision Transformer architecture operating on these supervertices, enabling anatomically informed learning from surface mesh representations. Compared with MSSM, MSSM+ identified more spatially extensive and statistically significant group differences between AD and CN. In AD versus CN classification, MSSM+ achieved a 3%p higher area under the precision–recall curve than MSSM. Vendor‐specific analyses further demonstrated reduced signal variability and consistently improved classification performance across MR manufacturers relative to CT, GWCs, and MSSM. These findings suggest that MSSM+ combined with SV‐ViT is a promising MRI‐based imaging marker for AD detection prior to CSF/PET confirmation. Keywords: Alzheimer's disease, gray‐to‐white matter contrast, imaging biomarker, MRI, multiscale structural mapping, supervertices We propose MSSM+, an extension of multiscale structural mapping (MSSM), together with surface supervertex mapping (SSVM) and a Supervertex Vision Transformer (SV‐ViT). Together, these methods exhibited better performance in detecting Alzheimer's disease and less variability across MR vendors than MSSM. 1. Introduction Alzheimer's disease (AD) is the most common cause of dementia worldwide, with neuropathological processes initiating years before the onset of clinical symptoms (Jack, Knopman, et al. 2010 ; Jack, Bernstein, Borowski, et al. 2010 ; Bateman et al. 2012 ). Detection is therefore critical for enabling preventive interventions and reducing disease burden. The National Institute on Aging and Alzheimer's Association (NIA–AA) research framework defines disease status along the amyloid/tau/neurodegeneration (A/T/N) scheme, in which structural magnetic resonance imaging (MRI) serves as a noninvasive indicator of N (Frisoni et al. 2010 ; Jack et al. 2018 ). Accordingly, structural MRI biomarkers such as hippocampal volume and cortical thickness (CT) (Fischl and Dale 2000 ) are used for screening prior to CSF/PET confirmation in research and clinical workflows. Plasma blood‐based biomarkers (BBMs), such as the phosphorylated‐tau 217/amyloid‐ β 42 ratio, have recently emerged. However, they lack anatomical specificity and have been validated primarily in adults 55 years and older, with significantly reduced accuracy in individuals aged 80 years and older (Ashton et al. 2024 ; Hansson et al. 2022 ; Janelidze et al. 2021 ; Cullen et al. 2023 ; Palmqvist et al. 2025 ; Hazan et al. 2025 ). To improve sensitivity, multiscale structural mapping (MSSM) (Jang et al. 2022 ) was introduced to integrate CT with gray‐to‐white matter contrasts (GWCs), and to project these high‐dimensional vertex‐wise features into a lower‐dimensional latent space using partial least squares‐discriminant analysis (PLS‐DA) (Lee et al. 2018 ). GWCs reflect microstructural boundary integrity and tissue contrast at the gray–white interface (Salat et al. 2009 , 2011 ; Westlye et al. 2010 ) and were implemented in MSSM as multiple depth‐wise intensity contrasts, which are sensitive to age‐ and disease‐related changes. Consistent with this view, a recent study reported that reduced GWCs are associated with markers of neurodegeneration, amyloid burden, and cognitive impairment across independent cohorts, supporting GWCs as a sensitive indicator of cortical tissue integrity (Xu et al. 2024 ). As a proof‐of‐concept, MSSM was initially evaluated on Siemens data, underscoring the need to establish cross‐vendor generalizability. Recent work further extended MSSM to derive structural indicators of β ‐amyloid neuropathology in preclinical AD, showing that MSSM signals are regionally altered and closely aligned with amyloid PET burden even when cortical thickness or hippocampal volume alone are not (Jang et al. 2024 ). Beyond CT and GWCs, additional vertex‐wise morphometric features may capture complementary aspects of cortical architecture relevant to AD. Sulcal depth (Sulc) indexes cortical folding and sulcal widening associated with tissue loss (Schaer et al. 2008 ; Thompson et al. 2004 ), whereas cortical curvature (Curv) quantifies local shape complexity and has been linked to cortical thinning and gyral collapse in neurodegeneration (Hogstrom et al. 2013 ). The white‐matter Jacobian determinant (Jacob) reflects localized expansion or contraction and has been used extensively in deformation‐based morphometry (Ashburner and Friston 2000 ; Gaser et al. 2001 ). The local gyrification index (LGI) quantifies folding extent and has been associated with developmental and pathological variation in cortical complexity (Schaer et al. 2008 ). Incorporating such intensity and geometry measures into a unified surface representation may enhance sensitivity to subtle cortical alterations. At the same time, it remains pivotal to identify compact feature subsets that maximize incremental value, as not all morphometric descriptors contribute equally in multivariate classification. In parallel with advances in feature design, deep learning methods have increasingly been applied to surface‐based neuroimaging data. Graph convolutional approaches such as Hypergraph Neural Network+ (HGNN+) (Gao et al. 2022 ), Spline‐based Graph CNN (SplineCNN) (Fey et al. 2018 ), SpiralNet++ (Gong et al. 2019 ), and Surface Vision Transformer (SiT) (Dahan et al. 2022 ) have been applied to cortical meshes. We developed surface supervertex mapping (SSVM) to extend this line of work by mapping cortical surfaces into nonoverlapping supervertices (SVs). Because cortical measurements reside on a non‐Euclidean two‐dimensional manifold embedded in three dimensions, projecting them onto planar grids distorts surface areas and neighborhood relations. Standard 2D encoders, which assume a Euclidean lattice, often require ad‐hoc unwrapping or resampling that can blur vertex‐wise signals. We therefore designed the Supervertex Vision Transformer (SV‐ViT) to directly use SVs as inputs to enable region‐of‐interest (ROI)‐informed learning. In this study, we extended MSSM by systematically evaluating vertex‐wise cortical features (GWCs, CT, Sulc, Curv, Jacob, and LGI) across multi‐vendor T1‐weighted datasets. We identified a compact subset, referred to as MSSM+, comprising GWCs, CT, Sulc, and Curv. We further proposed SSVM and designed SV‐ViT, benchmarking them against the latest models for AD versus cognitively normal (CN) classification. Together, these advances provide evidence for generalizability across Siemens, GE, and Philips scanners and support MSSM+ with SSVM and SV‐ViT as a promising surface‐based framework for MRI‐based detection of AD prior to CSF/PET confirmation in AD neurodegeneration. 2. Material and Methods 2.1. MRI Acquisition All data consisted of 3 T 3D sagittal T1‐weighted (T1w) acquisitions obtained from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database ( adni.loni.usc.edu ) (Jack et al. 2008 ; Jack, Bernstein, Borowski, et al. 2010 ) and Open Access Series of Imaging Studies (OASIS) (LaMontagne et al. 2019 ; Koenig et al. 2020 ) datasets. Data from ADNI–GO/2/3/4 and OASIS–3/4 were included. Accelerated images (factor = 2) were collected for ADNI, and fully sampled images were collected for OASIS to maximize sample size and capture data from diverse imaging protocols. Vendor‐specific implementations included Siemens MPRAGE, GE IR‐FSPGR and MPRAGE, and Philips MPRAGE (IR‐TFE). Representative parameters (TR, TE, TI, voxel size) by dataset and vendor were summarized in Table 1 . The reported TR values in the table followed vendor‐specific conventions and were not directly comparable across vendors. For Philips in ADNI–GO/2, TI was protocol‐based because converted headers sometimes reported 0. TABLE 1. T1w acquisition parameters by dataset, vendor, and protocol (3 T, sagittal, 3D). ADNI scans were accelerated, and OASIS scans were fully sampled. Values summarize representative acquisition parameters of the retained scans in the final analyzed dataset; ranges denote within‐phase variability. Dataset Vendor Protocol TR (ms) TE (ms) TI (ms) Voxel size (mm 3 ) ADNI–GO Siemens MPRAGE 2300 3.0 900 1.0 × 1.0 × 1.2 Philips MPRAGE 7 3.1 900 1.0 × 1.0 × 1.2 ADNI–2 Siemens MPRAGE 2300 3.0 900 1.0 × 1.0 × 1.2 GE IR‐FSPGR 7–8 2.8–3.2 400 1.0 × 1.0 × 1.2 Philips MPRAGE 7 3.1 900 1.0 × 1.0 × 1.2 ADNI–3 Siemens MPRAGE 2300 3.0 900 1.0 × 1.0 × 1.0 GE IR‐FSPGR 7–8 2.9–3.2 400 1.0 × 1.0 × 1.0 Philips MPRAGE 6–7 2.9–3.0 900 1.0 × 1.0 × 1.0 ADNI–4 Siemens MPRAGE 2300 2.9–3.2 900 1.0 × 1.0 × 1.0 GE IR‐FSPGR 7 2.9–3.1 400 1.0 × 1.0 × 1.0 GE MPRAGE 2300 3.0 900 1.0 × 1.0 × 1.0 Philips MPRAGE 7 2.9–3.0 900 1.0 × 1.0 × 1.0 OASIS–3 Siemens MPRAGE 2400 3.1–3.2 1000 1.0 × 1.0 × 1.0 OASIS–4 Siemens MPRAGE 2300 2.95 900 1.0 × 1.0 × 1.0 Open in a new tab 2.2. Participants We analyzed N = 1988 T1w scans from the datasets described above. After preprocessing, the final sample included 696 AD patients and 1292 CN individuals. Demographics and sample distributions by MR manufacturer (Siemens, GE, Philips) are summarized in Table 2 . All participants were unique, with no duplicate subjects retained. Unless otherwise specified, primary analyses used the large‐scale multi‐cohort datasets, and manufacturer‐specific analyses were conducted on subsets constructed by matching AD and CN counts across manufacturers to the smallest vendor (Philips). TABLE 2. Participant demographics ( n = 1988) by scanner vendors. Vendor Diagnosis n Age (year) Sex (F/M) Siemens AD 575 76.3 ± 7.6 274/301 CN 1123 69.8 ± 9.0 684/439 GE AD 69 75.3 ± 8.0 25/44 CN 103 73.2 ± 7.3 58/45 Philips AD 52 73.7 ± 7.7 25/27 CN 66 73.3 ± 8.2 43/23 Open in a new tab Note: Values are mean ± SD. All available participants were retained after preprocessing; no subjects were duplicated. 2.3. Cortical Surface Processing The overall MSSM+ analysis pipeline, including preprocessing, feature extraction, dimensionality reduction, age correction, group comparison, and classification, is summarized in Figure 1 . Cortical surfaces were reconstructed using FreeSurfer v7.4.1 (Fischl et al. 2004 ). All scans underwent standard FreeSurfer quality control; scans with significant reconstruction errors were excluded. Vertex‐wise cortical features (GWCs, CT, Sulc, Jacob, Curv, and LGI) were extracted for each participant and mapped to the fsaverage surface. Following the original MSSM procedure (Jang et al. 2022 ), GWCs were computed as multiple depth‐wise gray–white intensity contrasts. Gray matter intensities were sampled at 20%, 40%, 60%, and 80% cortical depth, and white matter intensities were sampled 0.5 and 1 mm beneath the gray–white boundary. Ratios of these gray and white matter intensities yielded eight vertex‐wise GWC maps per subject. This multiscale sampling quantified microstructural contrast at the cortical boundary, providing sensitivity to alterations associated with age and disease (Salat et al. 2009 , 2011 ; Westlye et al. 2010 ). FIGURE 1. Open in a new tab Overview of the MSSM+ procedure and classification using SSVM and SV‐ViT. AD, participants with Alzheimer's disease; CN, cognitively normal participants; CT, cortical thickness; Curv, cortical curvature; GM, gray matter; GWCs, gray‐to‐white matter contrasts; PLS‐DA, partial least squares‐discriminant analysis; Sulc, sulcal depth; SVs, supervertices; WM, white matter. 2.4. Dimensionality Reduction and Age Correction Because the features were multicollinear, PLS‐DA (Lee et al. 2018 ) was applied at each vertex to derive a latent component that maximally covaried with the diagnosis label (AD vs. CN). For a single vertex‐wise feature (e.g., CT alone), PLS‐DA was not needed. To reduce age‐related effects, vertex‐wise residualization (Zhang et al. 2023 ) was performed using polynomial regression models. Age modeling was conducted using an independent subset of 150 CN participants similar to previous work (Adaszewski et al. 2013 ; Tosun et al. 2016 ; Amoroso et al. 2018 ; Li et al. 2021 ). Specifically, we sampled 25 CN participants from each vendor‐by‐sex stratum (3 vendors × 2 sexes = 150). The residuals were used in classifications and statistical analyses. Data from these participants were not reused in any subsequent statistical analyses or classification studies. 2.5. Surface Supervertex Mapping We developed SSVM (Figure 2 ) to allow anatomically informed inter‐ and intra‐ROI attention and alleviate excessive computational burden by defining finer‐grained patches, or SVs. Vision Transformer (ViT) (Dosovitskiy et al. 2021 ) partitions 2D images into patches, which cannot directly represent cortical geometry. Surface Vision Transformer (SiT) (Dahan et al. 2022 ) adapts this concept using uniform icosahedral subdivision on the icosphere (ico6 → ico2) to produce equal‐area patches that preserve geometry but ignore anatomical boundaries. In contrast, SV‐ViT applied the proposed SSVM. The fsaverage icosphere (ico7) surface contains hundreds of thousands of triangular faces per hemisphere, making global self‐attention computationally prohibitive. Faces belonging to the medial wall were excluded before SV construction. A face was treated as a boundary face when its three vertices had mixed ROI labels. The boundary faces were excluded before breadth‐first search (BFS) initialization to prevent overlapping ROI assignment. We fixed the number of SVs to 1280 per hemisphere (the number of ico3 faces), ensuring balanced, topology‐preserving aggregation of cortical information. For each ROI, the number of patches was assigned in proportion to ROI size. Patch size, defined as the number of faces per patch, was chosen separately for each hemisphere as the largest feasible common value such that all patches in that hemisphere had the same number of faces. FIGURE 2. Open in a new tab Patch tokenization method (SSVM) for Supervertex Vision Transformer (SV‐ViT). 2.6. Supervertex Vision Transformer We designed SV‐ViT (Figure 3 ), which extended SiT by applying the proposed SSVM (Figure 2 ) to enable anatomically‐informed learning. SSVM processed cortical surface maps into SVs, which were then used as input to SV‐ViT, thereby enhancing efficiency and interpretability for surface‐based learning. The SVs were linearly embedded, and a learnable class (CLS) token of the same dimension was appended to the token sequence. A position embedding was added to each token. Then the tokens passed through stacked Transformer encoder blocks, each comprising a multi‐head self‐attention layer and a feed‐forward network. The attention mechanism captured relational dependencies among SVs, whereas position embedding preserved local cortical topography within SVs. From the output tokens, only the CLS token was fed to a multi‐layer perceptron (MLP) head to generate a logit for AD classification. FIGURE 3. Open in a new tab Architecture of the proposed Supervertex Vision Transformer (SV‐ViT). 2.7. AD versus CN Classification All the classification analyses followed a standard protocol. For all models, the inputs were the age‐corrected residuals. The tasks were formulated as binary classification (AD vs. CN) and evaluated using stratified four‐fold cross‐validation, with subject‐wise folds balanced by diagnosis. Within the training data, a further stratified split reserved 10% of subjects for validation. When multiple models or feature configurations were compared in the same subject subset, we reused identical folds and partitions across methods. The models were optimized with the AdamW optimizer and a cosine annealing learning rate scheduler to minimize the binary cross‐entropy loss. The classification performance was assessed on the held‐out test folds. The evaluation metrics were the area under the receiver operating characteristic curve (AUROC) and the area under the precision–recall curve (AUPRC). We used AUROC as the primary metric and AUPRC as the secondary metric for selecting MSSM+. We assessed all 63 combinations of vertex‐wise cortical features (GWCs, CT, Sulc, Jacob, Curv, and LGI) by conducting AD detection with SV‐ViT for each combination as input. We assessed the cross‐vendor robustness of MSSM+ across MR manufacturers by comparing its classification performance with CT, GWCs, and MSSM. Siemens, GE, and Philips scans were evaluated. Recent surface‐based models, including HGNN+ (Gao et al. 2022 ), SplineCNN (Fey et al. 2018 ), SpiralNet++ (Gong et al. 2019 ), and SiT (Dahan et al. 2022 ), were applied to AD versus CN classification with MSSM+ as input. Ablation studies were performed within the proposed framework. Incremental feature integration was examined by performing SV‐ViT for AD versus CN classification while adding the constituent MSSM+ features to the input one at a time in order of their relative importance. Leave‐one‐out analyses were then carried out by conducting SV‐ViT for AD versus CN classification, removing a single component (either an MSSM+ feature or the SV representation) at a time and keeping all remaining components unchanged. 2.8. Statistical Analysis Group‐level analyses were performed on age‐corrected residuals, as described above, to estimate the effect of AD diagnosis on CT, GWCs, MSSM, and MSSM+. Spatial smoothing was applied on the fsaverage surface with a Gaussian kernel of 10 mm full width at half maximum (FWHM) only for group‐level analyses. We used vertex‐wise General Linear Models (GLMs) (FreeSurfer v7.4.1). Per‐vertex p ‐values were reported on pial surface maps (called significance maps). Permutation‐based cluster‐wise correction for multiple comparisons was performed with 10,000 permutations, using a cluster‐forming threshold of 0.01 and a cluster‐wise corrected threshold of 0.05. Only the vertices that survived after correction were colored on the significance maps. Primary and manufacturer‐specific analyses were conducted. Differences in subject‐wise means of biomarkers (CT, GWCs, MSSM, and MSSM+) across MR manufacturers were investigated using Welch's ANOVA, which accommodates unequal variances and unequal sample sizes. Accordingly, this analysis was carried out on the entire dataset rather than on vendor‐specific subsets. When the omnibus Welch test was significant, pairwise Welch's t ‐tests between MR manufacturers (Siemens, GE, and Philips) were performed with Holm adjustment using a threshold of 0.05 for multiple comparisons. 3. Results 3.1. Effect of Extending MSSM with Additional Cortical Features We performed AD versus CN classification using SV‐ViT on large‐scale multi‐cohort datasets across all 63 combinations of cortical features (GWCs, CT, Sulc, Jacob, Curv, and LGI) to evaluate extensions of MSSM. Across all 63 feature combinations, MSSM+ achieved the highest AUROC and AUPRC (Table S1 ). As shown in Table 3 , MSSM+ (GWCs + CT + Sulc + Curv) achieved a mean AUROC of 0.93 and an AUPRC of 0.89, representing improvements of 1%p in AUROC and 3%p in AUPRC over MSSM. Among individual features, GWCs showed the strongest baseline, followed by CT, Sulc, Jacob, Curv, and LGI, in terms of AUROC and AUPRC (Figure 4 ). TABLE 3. AD versus CN classification performance for different cortical features in the combined multi‐vendor dataset. Features AUROC AUPRC CT (Fischl and Dale 2000 ) 0.8475 ± 0.0228 0.7402 ± 0.0113 GWCs (Salat et al. 2009 ) 0.8660 ± 0.0187 0.7653 ± 0.0232 MSSM (Jang et al. 2022 ) 0.9166 ± 0.0103 0.8588 ± 0.0186 MSSM+ 0.9312 ± 0.0051 0.8865 ± 0.0104 Open in a new tab Note: Values are reported as mean ± standard deviation across four‐fold cross‐validation on held‐out test folds. The best result for each performance metric was bolded. FIGURE 4. Open in a new tab AD versus CN classification performance of individual cortical features using SV‐ViT. Metrics were averaged across a four‐fold cross‐validation on held‐out test folds. Bars show AUROC and AUPRC for each feature: GWCs, CT, Sulc, Jacob, Curv, and LGI. Significance maps from the vertex‐wise group‐level analyses illustrated diagnosis‐related differences across CT, GWCs, MSSM, and MSSM+ (Figure 5 ). MSSM already showed broader and stronger effects than CT and GWCs. MSSM+ further extended these effects, with greater significance in the prefrontal and medial frontal cortices and more extensive significant regions in the paracentral cortex than in MSSM. FIGURE 5. Open in a new tab Significance maps from group‐level analysis of AD versus CN in the combined multi‐vendor dataset. 3.2. Generalizability across MR Manufacturers In vendor‐specific classifications, the AUROC and AUPRC of MSSM were higher and less variable across Siemens, GE, and Philips scanners than those of CT or GWCs. MSSM+ showed higher AUROC and AUPRC and lower variability across MR manufacturers than MSSM, achieving AUROC values of 0.93–0.94 across Siemens, GE, and Philips scanners (Table 4 ). TABLE 4. AD versus CN classification performance by MR manufacturer and features. Vendor‐specific subsets were down‐sampled to match the smallest vendor (Philips) in sample size and AD:CN ratio. Vendor Metric Features CT GWCs MSSM MSSM+ Siemens AUROC 0.8433 ± 0.0256 0.8356 ± 0.0272 0.9100 ± 0.0137 0.9307 ± 0.0124 AUPRC 0.7163 ± 0.0429 0.7398 ± 0.0424 0.8465 ± 0.0259 0.8847 ± 0.0246 GE AUROC 0.5664 ± 0.0997 0.8959 ± 0.0166 0.9172 ± 0.0169 0.9475 ± 0.0152 AUPRC 0.4920 ± 0.0824 0.8408 ± 0.0178 0.8743 ± 0.0363 0.9228 ± 0.0268 Philips AUROC 0.5892 ± 0.0831 0.9029 ± 0.0332 0.9142 ± 0.0247 0.9292 ± 0.0210 AUPRC 0.5838 ± 0.1117 0.8659 ± 0.0723 0.8745 ± 0.0653 0.8908 ± 0.0623 Open in a new tab Note: Values are reported as mean ± standard deviation across four‐fold cross‐validation on held‐out test folds. Vendor‐specific group‐level analyses were generally consistent with the classification results. For each manufacturer, the proportion of cortical vertices showing significant AD versus CN differences increased from CT to GWCs to MSSM and MSSM+, indicating more extensive significant regions. Variability in the proportion of significant vertices across Siemens, GE, and Philips differed by features, as summarized in Figure 6 . FIGURE 6. Open in a new tab Comparison of group‐level AD versus CN differences across MR manufacturers and features. Analyses of subject‐wise mean values using Welch's ANOVA showed significant manufacturer effects for CT, GWCs, MSSM, and MSSM+ (all Holm‐adjusted). Numbers above the brackets denote Holm‐adjusted p ‐values. Across measures, manufacturer‐related variability decreased from CT to GWCs, then to MSSM, and then to MSSM+ (Figure 7 ). FIGURE 7. Open in a new tab Cross‐vendor variability in subject‐wise mean values of CT, GWCs, MSSM, and MSSM+. 3.3. Performance of SV‐ViT with MSSM+ and Ablation Analyses In each hemisphere, approximately 2.9% of the 327,680 faces were classified as boundary faces and excluded before BFS initialization. Excluding the medial wall, no vertices were excluded at this stage because every vertex of an excluded boundary face was also shared by at least one retained neighboring face. Using MSSM+ as the standard input, we compared SV‐ViT with recent surface‐based models, including HGNN+, SplineCNN, SpiralNet++, and SiT. SV‐ViT and SiT, both Vision Transformer‐based architectures, yielded the largest AUROC and AUPRC, with SpiralNet++ and SplineCNN showing intermediate performance, and HGNN+ having lower mean values but the second‐smallest variability across folds. Among all models, SV‐ViT showed the highest mean AUROC and AUPRC and the smallest standard deviations (Table 5 ). TABLE 5. AD versus CN classification performance of recent surface‐based models using MSSM+ as input. Models AUROC AUPRC Hypergraph Neural Network+ (Gao et al. 2022 ) 0.8960 ± 0.0062 0.8133 ± 0.0128 Spline‐based Graph CNN (Fey et al. 2018 ) 0.9015 ± 0.0366 0.8430 ± 0.0427 SpiralNet++ (Gong et al. 2019 ) 0.9172 ± 0.0241 0.8575 ± 0.0412 Surface Vision Transformer (Dahan et al. 2022 ) 0.9299 ± 0.0161 0.8796 ± 0.0289 Supervertex Vision Transformer 0.9312 ± 0.0051 0.8865 ± 0.0104 Open in a new tab Note: Values are reported as mean ± standard deviation across four‐fold cross‐validation on held‐out test folds. The best result for each performance metric was bolded. Feature‐wise AD detection using SV‐ViT showed that GWCs had the highest AUROC and AUPRC among individual cortical features, followed by CT, Sulc, Jacob, Curv, and LGI (Figure 4 ). Incremental ablation for feature integration showed monotonic gains in AUROC and AUPRC as additional features were added to GWCs (Table 6 ). Adding CT improved the metrics, yielding MSSM (GWCs + CT), and subsequent inclusion of Sulc and Curv yielded smaller but consistent increases. TABLE 6. Incremental ablation of cortical feature sets for AD versus CN classification using SV‐ViT. Features AUROC AUPRC GWCs 0.8660 ± 0.0187 0.7653 ± 0.0232 GWCs + CT 0.9166 ± 0.0103 0.8588 ± 0.0186 GWCs + Sulc + CT 0.9230 ± 0.0066 0.8713 ± 0.0102 GWCs + Sulc + CT + Curv (MSSM+) 0.9312 ± 0.0051 0.8865 ± 0.0104 Open in a new tab Note: Values are reported as mean ± standard deviation across four‐fold cross‐validation on held‐out test folds. Leave‐one‐out ablations of MSSM+ further quantified the contribution of each component (Table 7 ). Removing GWCs yielded the largest reductions in metrics, followed by removing the SV representation, then CT, whereas removing Sulc or Curv produced smaller performance decreases. TABLE 7. Leave‐one‐out ablation of MSSM+ with SV‐ViT for AD versus CN classification. Each row reports performance when one component of the proposed framework (e.g., GWCs, CT, Sulc, Curv, or the SV representation) was removed. Δ AUROC and Δ AUPRC denote the absolute decrease in performance relative to the full framework. Components AUROC Δ AUROC AUPRC Δ AUPRC MSSM+ 0.9312 ± 0.0051 — 0.8865 ± 0.0104 — w/o Curv 0.9230 ± 0.0066 0.0082 0.8713 ± 0.0102 0.0153 w/o Sulc 0.9236 ± 0.0100 0.0076 0.8674 ± 0.0198 0.0192 w/o CT 0.9208 ± 0.0064 0.0105 0.8612 ± 0.0077 0.0253 w/o GWCs 0.8769 ± 0.0121 0.0543 0.7983 ± 0.0080 0.0882 w/o SV 0.8825 ± 0.0273 0.0487 0.8049 ± 0.0580 0.0816 Open in a new tab Note: Values are reported as mean ± standard deviation across four‐fold cross‐validation on held‐out test folds. 4. Discussion In this study, using large‐scale multi‐cohort datasets, we proposed MSSM+ and demonstrated that MSSM (GWCs + CT) could be improved by incorporating the additional cortical features Sulc and Curv. In addition, MSSM+ showed generalizability across MR manufacturers. We also designed SSVM and SV‐ViT and showed that anatomically informed learning on SVs could improve AD versus CN classification performance. MSSM+ (GWCs + CT + Sulc + Curv) yielded higher classification performance than MSSM, with approximately 3%p higher AUPRC and 1%p higher AUROC, as well as lower standard deviations across folds. In the group‐level analyses, MSSM+ also showed greater spatial extent and stronger statistical significance than MSSM. Orange ellipses highlight regions where MSSM+ showed significant effects in the superior, middle, and medial frontal cortices and more extensive regions in the paracentral cortices compared with MSSM (Figure 5 ). These findings suggest that extending the MSSM with Sulc and Curv yields measurable gains beyond CT and GWCs alone. These findings are also consistent with recent MSSM work suggesting that additional cortical surface features may further improve performance (Jang et al. 2024 ). Although Sulc and Curv showed smaller individual contributions than GWCs, they may provide complementary geometric information beyond tissue contrast and cortical thickness. The dominant contribution of GWCs is likewise consistent with prior evidence that GWC alterations are closely associated with neurodegeneration and cognitive decline across the AD continuum (Xu et al. 2024 ). In particular, these features characterize sulcal and local surface‐shape variation that may reflect cortical degeneration not fully captured by GWCs or CT alone (Schaer et al. 2008 ; Thompson et al. 2004 ; Hogstrom et al. 2013 ). Jacob showed higher single‐feature classification performance than Curv, and, except when added to the full MSSM+ feature combination, its inclusion generally increased performance. This pattern suggests that excessive feature complexity can degrade performance. Consistent with this, Jacob quantifies local tissue expansion and contraction and has been widely used in deformation‐based morphometry to detect subtle shape changes not captured by volumetric measures alone (Gaser et al. 2001 ; Ashburner and Friston 2000 ). LGI reduced classification performance when combined with other features, which we interpreted as indicating that LGI was not well suited to capture neurodegeneration; the fact that LGI has been reported to be primarily influenced by neurodevelopmental factors supports this interpretation (Schaer et al. 2008 ). Across vendor‐specific subsets, MSSM+ and MSSM showed similar AD versus CN classification performance. In group‐level analyses, both MSSM+ and MSSM identified more extensive regions of significant AD versus CN differences than CT and GWCs. Across vendors, variability in subject‐wise mean MSSM+ values, followed by MSSM, was lower than that of CT and GWCs, indicating that MSSM+ and MSSM were more consistent across MR manufacturers. SSVM enables inter‐ and intra‐ROI‐informed learning and alleviates computational complexity by defining finer‐grained patches, or SVs. The attention mechanism captures inter‐ROI dependencies among SVs. In contrast, the position embedding retains the intra‐ROI information within SVs. Increasing the attention dimension therefore increases computational cost while also increasing the capacity to capture inter‐regional dependencies. In our design, the attention dimension and position embedding dimension are constrained such that increasing one necessarily requires reducing the other. By controlling the number and size of SVs, our design optimizes an effective balance between computational burden and the richness of inter‐regional relations. SV‐ViT and SiT, ViT‐based models, exhibited higher classification performance than SpiralNet++, SplineCNN, and HGNN+, indicating that attention between cortical regions was pivotal for AD classification. The uniform icosahedral patches of SiT serve to partition the cortical surface into equally sized inputs for the model. In contrast, the SVs in SV‐ViT, in addition to serving this role, enable ROI‐informed learning. The SVs contribute to the performance improvement of SV‐ViT (Table 7 ). 4.1. Limitations For SSVM, we adopted a rule‐based, ROI‐constrained BFS to form equal‐face, edge‐connected, nonoverlapping supervertices within a single anatomical ROI. Only faces whose three vertices shared an identical anatomical ROI were included; all others were excluded to prevent ROI mixing. This yields a small, intentional under‐coverage concentrated at ROI borders and reflects design constraints rather than model failure. The vendor distribution was imbalanced (Siemens was the largest), raising the possibility that performance gains could be driven primarily by Siemens. Accordingly, we conducted vendor‐specific analyses, including classification (Table 4 ), group‐level analyses (Figure 6 ), and cross‐vendor variability analyses (Figure 7 ). These analyses showed that the advantages of MSSM+ were observed not only in the full sample (Table 3 ) but also consistently across all vendors. We also used PLS‐DA for vertex‐wise dimensionality reduction; while computationally efficient here, multi‐component or supervised variants may capture additional distributed effects. Age differences between groups are an inherent concern in AD studies. We mitigated this by fitting vertex‐wise polynomial age adjustment models on a small, balanced subset and applying the learned coefficients to the entire dataset; residual confounding may remain. 4.2. Future Directions Future directions include replacing the heuristic with a learned, connectivity‐preserving assignment that increases coverage while maintaining ROI fidelity and size balance in SSVM. This can be implemented on the cortical surface mesh adjacency graph using learned graph clustering instead of the current rule‐based BFS. A vertex‐based approach will prevent boundary faces consisting of vertices with mixed ROI labels and may improve ROI fidelity. We plan to refine architectures that more explicitly leverage cortical surface topology and vertex spatial relationships. Beyond the AD‐CN classification, this framework can be evaluated in mild cognitive impairment (MCI) to examine progression to AD, considering that prior work integrating structural features maps (Jang et al. 2022 ) demonstrated potential for longitudinal application. Further subgroup analyses by disease severity will also be important for better understanding stage‐specific patterns of disease‐related change. An additional priority is to integrate amyloid‐ β (A β ) and tau biomarkers from CSF or PET, thereby testing alignment with the A/T/N framework and evaluating prediction of A β ‐positive and tau‐positive status. We will further examine more refined age‐modeling strategies and conduct sensitivity analyses to assess the robustness of the age‐correction procedure. It will also be important to determine whether specific training‐sample compositions, including vendor‐balanced subsets, can further improve cross‐vendor performance. Scanner harmonization across MR manufacturers will be examined in future studies. Finally, extending the approach to multimodal structural MRI, such as T2‐weighted, fluid‐attenuated inversion recovery, and diffusion imaging, may further improve generalizability and sensitivity to early microstructural change. 5. Conclusion We proposed MSSM+, an extension of MSSM that integrates additional morphometric features at the vertex level. We also designed SSVM and SV‐ViT, which use SVs from SSVM to enable anatomically informed learning. Together, these approaches improved performance in AD versus CN classification for each MR vendor and yielded stronger, broader group contrast across large‐scale multi‐cohort datasets. These findings suggest that our approach may serve as a vendor‐generalizable and accessible MRI biomarker for the detection of AD prior to CSF/PET confirmation. Author Contributions Geonwoo Baek: conceptualization, methodology, software, formal analysis, writing – original draft. David H. Salat: supervision, validation, writing – review and editing. Ikbeom Jang: conceptualization, supervision, project administration, writing – review and editing. Funding This work was supported by the National Institutes of Health (NIH) grant (R21AG072431) of the U.S.; the National Research Foundation of Korea (NRF) grant funded by the Ministry of Science and ICT (MSIT) (RS‐2024‐00455720); the Korea National Institute of Health (KNIH) research projects (2024‐ER0407‐00, 2025‐ER0403‐00 and 2026‐ER0904‐00); the Korea Institute of Science and Technology Information with supercomputing resources including technical support (KSC‐2024‐CRE‐0021 and KSC‐2025‐CRE‐0065); and Hankuk University of Foreign Studies Research Fund of 2025. Ethics Statement The study protocols were approved by the Institutional Review Boards of the participating institutions. Consent Both ADNI and OASIS obtained informed consent from all participants. Conflicts of Interest The authors declare no conflicts of interest. Supporting information Table S1: AD vs. CN classification performance across all 63 combinations of cortical features. Filled circles (•) indicate included features (GWCs, CT, Sulc, Jacob, Curv, and LGI). For each combination, the highest and second‐highest values of each metric are highlighted in bold and underline respectively. Values are reported as mean ± standard deviation across four‐fold cross‐validation on held‐out test folds. HBM-47-e70548-s001.pdf (86.9KB, pdf) Acknowledgments Data collection and sharing for this project was funded by the Alzheimer's Disease Neuroimaging Initiative (ADNI) (National Institutes of Health Grant U01 AG024904) and DOD ADNI (Department of Defense award number W81XWH‐12‐2‐0012). ADNI is funded by the National Institute on Aging, the National Institute of Biomedical Imaging and Bioengineering, and through generous contributions from the following: AbbVie, Alzheimer's Association; Alzheimer's Drug Discovery Foundation; Araclon Biotech; BioClinica Inc.; Biogen; Bristol‐Myers Squibb Company; CereSpir Inc.; Cogstate; Eisai Inc.; Elan Pharmaceuticals Inc.; Eli Lilly and Company; EuroImmun; F. Hoffmann‐La Roche Ltd. and its affiliated company Genentech Inc.; Fujirebio; GE Healthcare; IXICO Ltd.; Janssen Alzheimer Immunotherapy Research & Development LLC.; Johnson & Johnson Pharmaceutical Research & Development LLC.; Lumosity; Lundbeck; Merck & Co. Inc.; Meso Scale Diagnostics LLC.; NeuroRx Research; Neurotrack Technologies; Novartis Pharmaceuticals Corporation; Pfizer Inc.; Piramal Imaging; Servier; Takeda Pharmaceutical Company; and Transition Therapeutics. The Canadian Institutes of Health Research is providing funds to support ADNI clinical sites in Canada. Private sector contributions are facilitated by the Foundation for the National Institutes of Health ( www.fnih.org ). The grantee organization is the Northern California Institute for Research and Education, and the study is coordinated by the Alzheimer's Therapeutic Research Institute at the University of Southern California. ADNI data are disseminated by the Laboratory for Neuro Imaging at the University of Southern California. OASIS data were provided by the OASIS project (LaMontagne et al. 2019 ; Koenig et al. 2020 ), Principal Investigators: D. Marcus, R. Buckner, J. Csernansky, and J. Morris. During the preparation of this work, the authors used ChatGPT (OpenAI) to improve readability and language clarity. The authors reviewed and edited all content, and take full responsibility for the content of the published article. Data Availability Statement Data used in the preparation of this article were obtained from the ADNI (Jack et al. 2008 ; Jack, Bernstein, Borowski, et al. 2010 ) and OASIS databases (LaMontagne et al. 2019 ; Koenig et al. 2020 ). The ADNI and OASIS data are available at https://adni.loni.usc.edu and https://www.oasis‐brains.org , respectively. As such, the investigators within the ADNI contributed to the design and implementation of ADNI and/or provided data but did not participate in analysis or writing of this report. A complete listing of ADNI investigators can be found at: http://adni.loni.usc.edu/wp‐content/uploads/how_to_apply/ADNI_Acknowledgement_List.pdf . The code used in this study is available at https://github.com/labhai/MSSMplus . References Adaszewski, S. , Dukart J., Kherif F., Frackowiak R., B. Draganski, and A.D.N.I . 2013. “How Early Can We Predict Alzheimer's Disease Using Computational Anatomy?” Neurobiology of Aging 34: 2815–2826. 10.1016/j.neurobiolaging.2013.06.015. [ DOI ] [ PubMed ] [ Google Scholar ] Amoroso, N. , Rocca M. L., Bellotti R., et al. 2018. “Alzheimer's Disease Diagnosis Based on the Hippocampal Unified Multi‐Atlas Network (HUMAN) Algorithm.” Biomedical Engineering Online 17: 6. 10.1186/s12938-018-0439-y. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Ashburner, J. , and Friston K.. 2000. “Voxel‐Based Morphometry—The Methods.” NeuroImage 11: 805–821. 10.1006/nimg.2000.0582. [ DOI ] [ PubMed ] [ Google Scholar ] Ashton, N. J. , Brum W. S., Di Molfetta G., et al. 2024. “Diagnostic Accuracy of a Plasma Phosphorylated Tau 217 Immunoassay for Alzheimer Disease Pathology.” JAMA Neurology 81: 255–263. 10.1001/jamaneurol.2023.5319. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Bateman, R. , Xiong C., Benzinger T., et al. 2012. “Clinical and Biomarker Changes in Dominantly Inherited Alzheimer's Disease.” New England Journal of Medicine 367: 795–804. 10.1056/NEJMoa1202753. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Cullen, N. C. , Janelidze S., Mattsson‐Carlgren N., et al. 2023. “Test–Retest Variability of Plasma Biomarkers in Alzheimer's Disease and Its Effects on Clinical Prediction Models.” Alzheimer's and Dementia 19: 797–806. 10.1002/alz.12706. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Dahan, S. , Fawaz A., Williams L. Z., et al. 2022. “Surface Vision Transformers: Attention‐Based Modelling Applied to Cortical Analysis.” in Proceedings of the International Conference on Medical Imaging With Deep Learning (MIDL) . PMLR. 282–303. https://proceedings.mlr.press/v172/dahan22a.html . Dosovitskiy, A. , Beyer L., Kolesnikov A., et al. 2021. “An Image Is Worth 16 × 16 Words: Transformers for Image Recognition at Scale.” in Proceedings of the International Conference on Learning Representations (ICLR) . https://openreview.net/forum?id=YicbFdNTTy . Fey, M. , Lenssen J. E., Weichert F., and Müller H.. 2018. “SplineCNN: Fast Geometric Deep Learning With Continuous B‐Spline Kernels.” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 869–877. 10.1109/CVPR.2018.00097. [ DOI ] Fischl, B. , and Dale A.. 2000. “Measuring the Thickness of the Human Cerebral Cortex From Magnetic Resonance Images.” Proceedings of the National Academy of Sciences of the United States of America 97: 11050–11055. 10.1073/pnas.200033797. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Fischl, B. , Kouwe A., and Destrieux C.. 2004. “Automatically Parcellating the Human Cerebral Cortex.” Cerebral Cortex 14: 11–22. 10.1093/cercor/bhg087. [ DOI ] [ PubMed ] [ Google Scholar ] Frisoni, G. , Fox N., Jack C., Scheltens P., and Thompson P.. 2010. “The Clinical Use of Structural MRI in Alzheimer's Disease.” Nature Reviews. Neurology 6: 67–77. 10.1038/nrneurol.2009.215. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Gao, Y. , Feng Y., Ji S., and Ji R.. 2022. “HGNN+: General Hypergraph Neural Networks.” IEEE Transactions on Pattern Analysis and Machine Intelligence 45: 3181–3199. 10.1109/TPAMI.2022.3182052. [ DOI ] [ PubMed ] [ Google Scholar ] Gaser, C. , Nenadic I., Buchsbaum B., Hazlett E., and Buchsbaum M.. 2001. “Deformation‐Based Morphometry and Its Relation to Conventional Volumetry of Brain Lateral Ventricles in MRI.” NeuroImage 13: 1140–1145. 10.1006/nimg.2001.0771. [ DOI ] [ PubMed ] [ Google Scholar ] Gong, S. , Chen L., Bronstein M. M., and Zafeiriou S.. 2019. “SpiralNet++: A Fast and Highly Efficient Mesh Convolution Operator.” in Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops (ICCVW) (IEEE). 4141–4148. 10.1109/ICCVW.2019.00509. [ DOI ] Hansson, O. , Edelmayer R. M., Boxer A. L., et al. 2022. “The Alzheimer's Association Appropriate Use Recommendations for Blood Biomarkers in Alzheimer's Disease.” Alzheimer's and Dementia 18: 2669–2686. 10.1002/alz.12756. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Hazan, J. , Abel E., Grilo M. R., et al. 2025. “How Well Do Plasma Alzheimer's Disease Biomarkers Reflect the CSF Amyloid Status?” Journal of Neurology, Neurosurgery and Psychiatry 96: 566–572. 10.1136/jnnp-2024-334122. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Hogstrom, L. , Westlye L., Walhovd K., and Fjell A.. 2013. “The Structure of the Cerebral Cortex Across Adult Life: Age‐Related Patterns of Surface Area, Thickness, and Gyrification.” Cerebral Cortex 23: 2521–2530. 10.1093/cercor/bhs231. [ DOI ] [ PubMed ] [ Google Scholar ] Jack, C. , Bennett D., Blennow K., et al. 2018. “Nia‐Aa Research Framework: Toward a Biological Definition of Alzheimer's Disease.” Alzheimer's and Dementia 14: 535–562. 10.1016/j.jalz.2018.02.018. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Jack, C. , Knopman D., Jagust W., et al. 2010. “Hypothetical Model of Dynamic Biomarkers of the Alzheimer's Pathological Cascade.” Lancet Neurology 9: 119–128. 10.1016/S1474-4422(09)70299-6. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Jack, C. R. J. , Bernstein M. A., Borowski B. J., et al. 2010. “Update on the Magnetic Resonance Imaging Core of the Alzheimer's Disease Neuroimaging Initiative.” Alzheimer's and Dementia 6: 212–220. 10.1016/j.jalz.2010.03.004. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Jack, C. R. J. , Bernstein M. A., Fox N. C., et al. 2008. “The Alzheimer's Disease Neuroimaging Initiative (ADNI): MRI Methods.” Journal of Magnetic Resonance Imaging 27: 685–691. 10.1002/jmri.21049. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Janelidze, S. , Teunissen C. E., Zetterberg H., et al. 2021. “Head‐To‐Head Comparison of 8 Plasma Amyloid‐ β 42/40 Assays in Alzheimer Disease.” JAMA Neurology 78: 1375–1382. 10.1001/jamaneurol.2021.3180. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Jang, I. , Li B., Rashid B., et al. 2024. “Brain Structural Indicators of β ‐Amyloid Neuropathology.” Neurobiology of Aging 136: 157–170. 10.1016/j.neurobiolaging.2024.01.005. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Jang, I. , Li B., Riphagen J. M., Dickerson B. C., and Salat D. H.. 2022. “Multiscale Structural Mapping of Alzheimer's Disease Neurodegeneration.” NeuroImage: Clinical 33: 102948. 10.1016/j.nicl.2022.102948. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Koenig, L. N. , Day G. S., Salter A., et al. 2020. “Select Atrophied Regions in Alzheimer Disease (Sara): An Improved Volumetric Model for Identifying Alzheimer Disease Dementia.” NeuroImage: Clinical 26: 102248. 10.1016/j.nicl.2020.102248. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] LaMontagne, P. J. , Benzinger T. L. S., Morris J. C., et al. 2019. “Oasis‐3: Longitudinal Neuroimaging, Clinical, and Cognitive Dataset for Normal Aging and Alzheimer Disease.” medRxiv. 10.1101/2019.12.13.19014902. [ DOI ] [ Google Scholar ] Lee, L. C. , Liong C. Y., and Jemain A. A.. 2018. “Partial Least Squares‐Discriminant Analysis (PLS‐DA) for Classification of High‐Dimensional (HD) Data: A Review of Contemporary Practice Strategies and Knowledge Gaps.” Analyst 143: 3526–3539. 10.1039/c8an00599k. [ DOI ] [ PubMed ] [ Google Scholar ] Li, B. Y. , Jang I., Riphagen J., et al. 2021. “Identifying Individuals With Alzheimer's Disease‐Like Brains Based on Structural Imaging in the Human Connectome Project Aging Cohort.” Human Brain Mapping 42: 5535–5546. 10.1002/hbm.25626. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Palmqvist, S. , Warmenhoven N., Anastasi F., et al. 2025. “Plasma Phospho‐tau217 for Alzheimer's Disease Diagnosis in Primary and Secondary Care Using a Fully Automated Platform.” Nature Medicine 31: 2036–2043. 10.1038/s41591-025-03622-w. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Salat, D. , Lee S., van der Kouwe A., et al. 2009. “Age‐Associated Alterations in Cortical Gray and White Matter Signal Intensity and Gray to White Matter Contrast.” NeuroImage 48: 21–28. 10.1016/j.neuroimage.2009.06.074. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Salat, D. H. , Chen J. J., van der Kouwe A. J., Greve D. N., Fischl B., and Rosas H. D.. 2011. “Hippocampal Degeneration Is Associated With Temporal and Limbic Gray Matter/White Matter Tissue Contrast in Alzheimer's Disease.” NeuroImage 54: 1795–1802. 10.1016/j.neuroimage.2010.10.034. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Schaer, M. , Cuadra M., Tamarit L., et al. 2008. “A Surface‐Based Approach to Quantify Local Cortical Gyrification.” IEEE Transactions on Medical Imaging 27: 161–170. 10.1109/TMI.2007.903576. [ DOI ] [ PubMed ] [ Google Scholar ] Thompson, P. , Hayashi K., de Zubicaray G., et al. 2004. “Mapping Cortical Change in Alzheimer's Disease, Brain Development, and Schizophrenia.” NeuroImage 23: S2–S18. 10.1016/j.neuroimage.2004.07.071. [ DOI ] [ PubMed ] [ Google Scholar ] Tosun, D. , Chen Y. F., Yu P., et al. 2016. “Amyloid Status Imputed From a Multimodal Classifier Including Structural MRI Distinguishes Progressors From Nonprogressors in a Mild Alzheimer's Disease Clinical Trial Cohort.” Alzheimer's and Dementia 12: 977–986. 10.1016/j.jalz.2016.03.009. [ DOI ] [ PubMed ] [ Google Scholar ] Westlye, L. , Walhovd K., Dale A., et al. 2010. “Life‐Span Changes of the Human Brain White Matter: Diffusion Tensor Imaging (DTI) and Volumetry.” Cerebral Cortex 20: 2055–2068. 10.1093/cercor/bhp280. [ DOI ] [ PubMed ] [ Google Scholar ] Xu, X. , Jang I., Zhang J., et al. 2024. “Cortical Gray to White Matter Signal Intensity Ratio as a Sign of Neurodegeneration and Cognition Independent of β ‐Amyloid in Dementia.” Human Brain Mapping 45: e26532. 10.1002/hbm.26532. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Zhang, B. , Zhang S., Feng J., and Zhang S.. 2023. “Age‐Level Bias Correction in Brain Age Prediction.” NeuroImage: Clinical 37: 103319. 10.1016/j.nicl.2023.103319. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials Table S1: AD vs. CN classification performance across all 63 combinations of cortical features. Filled circles (•) indicate included features (GWCs, CT, Sulc, Jacob, Curv, and LGI). For each combination, the highest and second‐highest values of each metric are highlighted in bold and underline respectively. Values are reported as mean ± standard deviation across four‐fold cross‐validation on held‐out test folds. HBM-47-e70548-s001.pdf (86.9KB, pdf) Data Availability Statement Data used in the preparation of this article were obtained from the ADNI (Jack et al. 2008 ; Jack, Bernstein, Borowski, et al. 2010 ) and OASIS databases (LaMontagne et al. 2019 ; Koenig et al. 2020 ). The ADNI and OASIS data are available at https://adni.loni.usc.edu and https://www.oasis‐brains.org , respectively. As such, the investigators within the ADNI contributed to the design and implementation of ADNI and/or provided data but did not participate in analysis or writing of this report. A complete listing of ADNI investigators can be found at: http://adni.loni.usc.edu/wp‐content/uploads/how_to_apply/ADNI_Acknowledgement_List.pdf . The code used in this study is available at https://github.com/labhai/MSSMplus . Articles from Human Brain Mapping are provided here courtesy of Wiley ACTIONS View on publisher site PDF (3.9 MB) Cite Collections Permalink RESOURCES Similar articles Cited by other articles Links to NCBI Databases Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top

Related documents

• Molecular Imaging Modality by Positron Emission Tomography Using 18F-X : Study of Microglial Activation in Amyotrophic Lateral Sclerosis#851128
• Spatial proteomics of Alzheimer's disease-specific human microglial states.#9213
• Locus coeruleus volume and cell population changes during Alzheimer's disease progression: A stereological study in human postmortem brains with potential implication for early‐stage biomarker discovery#1174887
• The contribution of oxidative stress in the protein damage and DNA lesion in Alzheimer's disease neuropathology#957884
• Oral Fecal Microbiota Transplant Feasibility Study in Alzheimer's Disease#981078