A Hybrid Vision Transformer for T1-Weighted MRI-Based Alzheimer's Disease Staging with Biomarker Fusion
Engineering, Technology & Applied Science Research · 04.04.2026 · 16 (2) · 34546-34552
DOI: 10.48084/etasr.17151 ↗Kurz gefasst
Fachartikel in der Zeitschrift Engineering, Technology & Applied Science Research, online erschienen am 04.04.2026. Verfasst von Sonali Deshpande und 1 weiteren Autor:innen. Die Arbeit wurde bisher 1-mal zitiert (Crossref) und nennt 25 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Themen: Demenz
Abstract (Originalsprache)
Magnetic Resonance Imaging (MRI) with deep learning is widely applied for computer-aided diagnosis of Alzheimer's Disease (AD); however, existing models generally struggle with small and imbalanced datasets and often fail to thoroughly utilize anatomically meaningful biomarkers, which are essential for the detection of AD in the early stages. In this work, we tackle these shortcomings by introducing Hybrid Transformer for Alzheimer's Diagnosis (HyTraAD), a hybrid transformer-based model for 4-stage AD classification, including Cognitively Normal (CN), Early Mild Cognitive Impairment (EMCI), Mild Cognitive Impairment (MCI), and AD, from T1-weighted structural MRIs. Our approach combines a Residual Network (ResNet) 50 feature extractor with a light-weight Vision Transformer (ViT) encoder and directly fuses three volumetric biomarkers: hippocampal volume, temporal parietal cortical thickness, and ventricular volume into the learned representation. To address dataset imbalance and improve robustness, a noise-tolerant preprocessing pipeline is introduced, combining Tomek Links for removing borderline samples with the Synthetic Minority Over-sampling Technique (SMOTE) for balancing underrepresented classes. The model was evaluated on the Alzheimer's Disease Neuroimaging Initiative (ADNI) cohort consisting of 1,850 subjects. Experimental results demonstrate that HyTraAD achieves 99.81% overall accuracy, a macro F1-score of 0.99, and an EMCI recall of 0.95 on the test set, outperforming recent hybrid Convolutional Neural Network (CNN)-ViT architectures such as Hybrid ResNet-50 + ViT (RViT) and Visual Geometry Group (VGG)-based TSwinformer. Ablation studies further confirm that both biomarker integration and the proposed Tomek Links-SMOTE preprocessing strategy contribute significantly to the performance improvements, particularly in enhancing sensitivity to EMCI cases. Collectively, the results demonstrate that HyTraAD provides a flexible and interpretable framework for MRI-based staging of AD, which is promising for future deployment in clinically oriented decision-support systems, particularly in multi-center and multimodal diagnostic settings.
Zitieren
APA
Deshpande, S., & Kulkarni, N. (2026). A Hybrid Vision Transformer for T1-Weighted MRI-Based Alzheimer's Disease Staging with Biomarker Fusion. Engineering, Technology & Applied Science Research, 16(2), 34546-34552. https://doi.org/10.48084/etasr.17151
Vancouver
Deshpande S, Kulkarni N. A Hybrid Vision Transformer for T1-Weighted MRI-Based Alzheimer's Disease Staging with Biomarker Fusion. Engineering, Technology & Applied Science Research. 2026;16(2):34546-34552. doi:10.48084/etasr.17151
BibTeX
@article{Deshpande2026,
author = {Deshpande, Sonali and Kulkarni, Nilima},
title = {A Hybrid Vision Transformer for T1-Weighted MRI-Based Alzheimer's Disease Staging with Biomarker Fusion},
journal = {Engineering, Technology & Applied Science Research},
year = {2026},
volume = {16},
number = {2},
pages = {34546--34552},
doi = {10.48084/etasr.17151}
}
EUVIMED