CoChat

Differential diagnosis of frontotemporal dementia subtypes with explainable deep learning on structural MRI

Open access
Ma, Da, Stocks, Jane, Rosen, Howard, Kantarci, Kejal, Lockhart, Samuel N., Bateman, James R., Craft, Suzanne, Gurcan, Metin N., Popuri, Karteek, Beg, Mirza Faisal, Wang, Lei, on behalf of the ALLFTD consortium · 2024 · Frontiers in Neuroscience

The DOI matches its Crossref record; 70 of the 71 references it cites were found in a registry or an index; 19 of the 26 sources its claims cite were checked for support.

Abstract

Background: Frontotemporal dementia (FTD) represents a collection of neurobehavioral and neurocognitive syndromes that are associated with a significant degree of clinical, pathological, and genetic heterogeneity. Such heterogeneity hinders the identification of effective biomarkers, preventing effective targeted recruitment of participants in clinical trials for developing potential interventions and treatments. In the present study, we aim to automatically differentiate patients with three clinical phenotypes of FTD, behavioral-variant FTD (bvFTD), semantic variant PPA (svPPA), and nonfluent variant PPA (nfvPPA), based on their structural MRI by training a deep neural network (DNN). Methods: Data from 277 FTD patients (173 bvFTD, 63 nfvPPA, and 41 svPPA) recruited from two multi-site neuroimaging datasets: the Frontotemporal Lobar Degeneration Neuroimaging Initiative and the ARTFL-LEFFTDS Longitudinal Frontotemporal Lobar Degeneration databases. Raw T1-weighted MRI data were preprocessed and parcellated into patch-based ROIs, with cortical thickness and volume features extracted and harmonized to control the confounding effects of sex, age, total intracranial volume, cohort, and scanner difference. A multi-type parallel feature embedding framework was trained to classify three FTD subtypes with a weighted cross-entropy loss function used to account for unbalanced sample sizes. Feature visualization was achieved through post-hoc analysis using an integrated gradient approach. Results: The proposed differential diagnosis framework achieved a mean balanced accuracy of 0.80 for bvFTD, 0.82 for nfvPPA, 0.89 for svPPA, and an overall balanced accuracy of 0.84. Feature importance maps showed more localized differential patterns among different FTD subtypes compared to groupwise statistical mapping. Conclusion: In this study, we demonstrated the efficiency and effectiveness of using explainable deep-learning-based parallel feature embedding and visualization framework on MRI-derived multi-type structural patterns to differentiate three clinically defined subphenotypes of FTD: bvFTD, nfvPPA, and svPPA, which could help with the identification of at-risk populations for early and precise diagnosis for intervention planning.

Overview

CoChat Summary

## TL;DR

This study trained an explainable deep neural network on structural T1-weighted MRI to differentiate three clinical frontotemporal dementia subtypes: bvFTD, nfvPPA, and svPPA. Using 277 patients from two multi-site cohorts, the model achieved overall balanced accuracy of 0.84 and produced feature-importance maps that aligned with known subtype-specific atrophy patterns.

## Key findings

  • The proposed framework achieved mean balanced accuracy of 0.80 for bvFTD, 0.82 for nfvPPA, 0.89 for svPPA, and 0.84 overall.
  • In the more detailed results, the best model achieved 79.7% balanced accuracy for bvFTD, 81.9% for nfvPPA, 89.2% for svPPA, and 83.6% overall.
  • In the ablation study, thickness-only features outperformed volume-only features for bvFTD, svPPA, and overall performance; simple concatenation of thickness and volume reduced performance relative to thickness-only, while the proposed multi-level parallel embedding improved performance for bvFTD, nfvPPA, and overall.
  • Integrated Gradient maps showed subtype-relevant regions: left temporal patches for nfvPPA and svPPA, inferior frontal/frontal operculum regions for nfvPPA, and left anterior temporal, frontal opercular/insular, and bilateral frontal pole regions for bvFTD.

## Methods

  • Design: Retrospective multi-class classification study using an explainable deep-learning pipeline for differential diagnosis of FTD subtypes from structural MRI.
  • Sample/data: 277 FTD patients total: 173 bvFTD, 63 nfvPPA, 41 svPPA, drawn from the ALLFTD and FTLDNI/NIFD multi-site cohorts; cognitively normal controls were excluded because only 27 were available.
  • Main analysis: Raw T1 MRI were parcellated into 360 cortical patches plus 15 subcortical structures for 735 features total; features were harmonized with a regression-based w-score approach controlling for cohort, scanner/coil, sex, age, and TIV, then classified with a two-level MLP-based parallel feature embedding DNN evaluated by 10-fold nested cross-validation and interpreted with Integrated Gradients.

## Limitations

  • The paper states that it did not include cognitively normal controls because the sample was too small; it therefore used bvFTD as the reference group for harmonization, which the authors note means the feature-importance maps mainly reflect differences among the three FTD subtypes rather than differences versus healthy aging.
  • The authors state that they treated demographic variables as confounders, but future work could incorporate demographic information directly into the model; this indicates the current framework may miss useful demographic signal.
  • Evident from the text: the study used only structural T1-weighted MRI and only three clinical subtypes, so generalization to other modalities or broader dementia groups was not tested in this paper.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.3389/fnins.2024.1331677

Cites

70 of 71 references verified1 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1 Schematic diagram of the framework in this study, which consists of: (1) feature extraction to derive patch-based multi-type features of cortical thickness and cortical/subcortical volumes; (2) W-score-based feature harmonization to control confounding factors such as scanner difference and study site bias, as well as demographic-related covariates; (3) the differential diagnosis model using multi-layer-perceptron (MLP)-based multi-level parallel feature embedding deep neural network to achieve FTD subtype classification; (4) neuroimaging-derived brain differential patterns among different FTD subtypes.
Figure 1 Schematic diagram of the framework in this study, which consists of: (1) feature extraction to derive patch-based multi-type features of cortical thickness and cortical/subcortical volumes; (2) W-score-based feature harmonization to control confounding factors such as scanner difference and study site bias, as well as demographic-related covariates; (3) the differential diagnosis model using multi-layer-perceptron (MLP)-based multi-level parallel feature embedding deep neural network to achieve FTD subtype classification; (4) neuroimaging-derived brain differential patterns among different FTD subtypes.CC BY · PubMed Central
Figure 2 The schematic diagram of the neural network architecture of the multi-level parallel feature embedding framework used in this study to achieve accurate classification to differentiate FTD subtypes. Each block represents a multi-layer-perceptron (MLP) block. The number displayed in each of the MLP-based feature embedding blocks indicates the number of nodes in the corresponding layer.
Figure 2 The schematic diagram of the neural network architecture of the multi-level parallel feature embedding framework used in this study to achieve accurate classification to differentiate FTD subtypes. Each block represents a multi-layer-perceptron (MLP) block. The number displayed in each of the MLP-based feature embedding blocks indicates the number of nodes in the corresponding layer.CC BY · PubMed Central
Figure 3 Effects of feature harmonization in preprocessings. The paranomic heatmap shows the Z-scores of thickness and volume features before (left) and after (right) the data harmonization. Z-score values for each features represents the difference between individual measurements compared to the reference group mean, standardized by the reference group standard deviation. Negative Z-scores indicate lower value than the reference mean (i.e., smaller volume, thinner cortex), while positive Z-scores represent higher value than reference mean (i.e., larger volume, thicker cortex). Cohort-dependent biases were noticeable before the harmonization (left), which were reduced after the GLM-based feature harmonization step (right).
Figure 3 Effects of feature harmonization in preprocessings. The paranomic heatmap shows the Z-scores of thickness and volume features before (left) and after (right) the data harmonization. Z-score values for each features represents the difference between individual measurements compared to the reference group mean, standardized by the reference group standard deviation. Negative Z-scores indicate lower value than the reference mean (i.e., smaller volume, thinner cortex), while positive Z-scores represent higher value than reference mean (i.e., larger volume, thicker cortex). Cohort-dependent biases were noticeable before the harmonization (left), which were reduced after the GLM-based feature harmonization step (right).CC BY · PubMed Central
Figure 4 Classification performance in terms of balanced accuracy comparison among different combination of input features and network architecture design: (1) volume-only input features; (2) thickness-only input features; (3) Joint volume and thickness input feature using single-level MLP; and (4) Joint volume and thickness input feature using two-level MLP. The balanced accuracy for each individual FTD subtype (bvFTD, nfvPPA, and svPPA) as well as overall balance accuracy was reported. The black diamond box both in each subtype group as well as the overall performance represent the values of balanced accuracy that beyond two standard deviations among the 10-fold cross-validation.
Figure 4 Classification performance in terms of balanced accuracy comparison among different combination of input features and network architecture design: (1) volume-only input features; (2) thickness-only input features; (3) Joint volume and thickness input feature using single-level MLP; and (4) Joint volume and thickness input feature using two-level MLP. The balanced accuracy for each individual FTD subtype (bvFTD, nfvPPA, and svPPA) as well as overall balance accuracy was reported. The black diamond box both in each subtype group as well as the overall performance represent the values of balanced accuracy that beyond two standard deviations among the 10-fold cross-validation.CC BY · PubMed Central
Figure 5 Differential cortical patterns for each of the FTD subtype. The cortical manifold plot visualizes populational average feature importance map using Integrated Gradient based feature importance analysis projected onto the template cortical manifold (HCP-MMP1 atlas) for both the cortical thickness (left) and volume (right) features. The color maps represent Integrated Gradient based feature importance scores ranging from −0.06 to 0.06. The magnitude of the feature attribution (i.e., absolute value) represent the influence of each feature towards the output classification, while the sign of the feature attribution (i.e., positive and negative) reflect the direction of the feature influence towards the classification output. Attributions that are close to zero represent features that have minimal influence in models prediction.
Figure 5 Differential cortical patterns for each of the FTD subtype. The cortical manifold plot visualizes populational average feature importance map using Integrated Gradient based feature importance analysis projected onto the template cortical manifold (HCP-MMP1 atlas) for both the cortical thickness (left) and volume (right) features. The color maps represent Integrated Gradient based feature importance scores ranging from −0.06 to 0.06. The magnitude of the feature attribution (i.e., absolute value) represent the influence of each feature towards the output classification, while the sign of the feature attribution (i.e., positive and negative) reflect the direction of the feature influence towards the classification output. Attributions that are close to zero represent features that have minimal influence in models prediction.CC BY · PubMed Central
Figure 6 the statistical cortical mapping for each FTD subtype in which the patch-wise cortical features (both thickness and volume) were statistically compared with the combination of remaining populations that belong to the combination of the other two FTD subtypes.
Figure 6 the statistical cortical mapping for each FTD subtype in which the patch-wise cortical features (both thickness and volume) were statistically compared with the combination of remaining populations that belong to the combination of the other two FTD subtypes.CC BY · PubMed Central

Verification

Claims the paper makes

14 Valid · 0 Not valid · 5 Not checked · 32 Own results and interpretations

22 more claims

Create a free account to see all 51 claims and to verify the 2 references still open.

See all claims — free account
  • Probably not a substantial claim. 12
  • It may state two separate claims as one. 14
  • The quote only partly states this claim. 1

Create a free account to see these claims.

Citations

Cites

Show all 71 references

Cited by

See all 9

Metadata

Authors
Ma, D., Stocks, J., Rosen, H., Kantarci, K., Lockhart, S. N., Bateman, J. R., Craft, S., Gurcan, M. N., Popuri, K., Beg, M. F., Wang, L., & on behalf of the ALLFTD consortium
Affiliations
  • Wake Forest University Health Sciences (US)
  • Northwestern University (US)
  • University of California, San Francisco (US)
  • Mayo Clinic (US)
  • Memorial University of Newfoundland (CA)
  • Simon Fraser University (CA)
Corresponding author
  • Ma, Da
  • Stocks, Jane
  • Wang, Lei
Venue
Frontiers in Neuroscience
Published
2024-02-07
Volume
18
Pages
1331677-1331677
ISSN
1662-453X (ISSN-L)
OpenAlex
W4391609079
Cited by
9
References
71
Citations by year
2026: 5 · 2025: 3 · 2024: 1
FWCI
2.63
Type
journal article
Language
en
Open access
Locations
  • Frontiers in Neuroscience (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • eScholarship (California Digital Library) (submittedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
National Institute on Aging (R01AG055121, R56AG055121, T32AG020506, P30AG072977, P30AG072947)
Topics
  • Health Sciences › Medicine › Psychiatry and Mental health › Dementia and Cognitive Impairment Research
  • Life Sciences › Neuroscience › Cellular and Molecular Neuroscience › Alzheimer's disease research and treatments
  • Health Sciences › Medicine › Clinical Neurology › Amyotrophic Lateral Sclerosis Research
Keywords
frontotemporal dementia, differential diagnosis, behavioral variant FTD, semantic variant PPA, structural MRI, explainable deep learning, cortical thickness, cortical volume, integrated gradients, feature visualization, biomarkers

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol and qwen3.8-omni-flash).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Differential diagnosis of frontotemporal dementia subtypes with explainable deep learning on structural MRI. https://app.cochat.ai/papers/10.3389/fnins.2024.1331677
The paper
Ma, D., Stocks, J., Rosen, H., Kantarci, K., Lockhart, S. N., Bateman, J. R., Craft, S., Gurcan, M. N., Popuri, K., Beg, M. F., Wang, L., & on behalf of the ALLFTD consortium. (2024). Differential diagnosis of frontotemporal dementia subtypes with explainable deep learning on structural MRI. Frontiers in Neuroscience. https://doi.org/10.3389/fnins.2024.1331677

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.