Open Access

ARTICLE

DA-ViT: Deformable Attention Vision Transformer for Alzheimer’s Disease Classification from MRI Scans

Abdullah G. M. Almansour^1,*, Faisal Alshomrani², Abdulaziz T. M. Almutairi³, Easa Alalwany⁴, Mohammed S. Alshuhri¹, Hussein Alshaari⁵, Abdullah Alfahaid⁴

1 Radiology and Medical Imaging Department, College of Applied Medical Sciences, Prince Sattam Bin Abdulaziz University, Alkharj, 11942, Saudi Arabia
2 Department of Diagnostic Radiology Technology, College of Applied Medical Science, Taibah University, Medinah, 41477, Saudi Arabia
3 Department of Computer, College of Science and Humanities, Shaqra University, Shaqra, 11961, Saudi Arabia
4 Department of Computer Science, College of Computer Science and Engineering, Taibah University, Yanbu, 46522, Saudi Arabia
5 College of Applied Medical Sciences, Radiological Sciences Department, Najran University, Najran, 61441, Saudi Arabia

* Corresponding Author: Abdullah G. M. Almansour. Email:

(This article belongs to the Special Issue: Intelligent Medical Decision Support Systems: Methods and Applications)

Computer Modeling in Engineering & Sciences 2025, 144(2), 2395-2418. https://doi.org/10.32604/cmes.2025.069661

Received 27 June 2025; Accepted 12 August 2025; Issue published 31 August 2025

Abstract

The early and precise identification of Alzheimer’s Disease (AD) continues to pose considerable clinical difficulty due to subtle structural alterations and overlapping symptoms across the disease phases. This study presents a novel Deformable Attention Vision Transformer (DA-ViT) architecture that integrates deformable Multi-Head Self-Attention (MHSA) with a Multi-Layer Perceptron (MLP) block for efficient classification of Alzheimer’s disease (AD) using Magnetic resonance imaging (MRI) scans. In contrast to traditional vision transformers, our deformable MHSA module preferentially concentrates on spatially pertinent patches through learned offset predictions, markedly diminishing processing demands while improving localized feature representation. DA-ViT contains only 0.93 million parameters, making it exceptionally suitable for implementation in resource-limited settings. We evaluate the model using a class-imbalanced Alzheimer’s MRI dataset comprising 6400 images across four categories, achieving a test accuracy of 80.31%, a macro F1-score of 0.80, and an area under the receiver operating characteristic curve (AUC) of 1.00 for the Mild Demented category. Thorough ablation studies validate the ideal configuration of transformer depth, headcount, and embedding dimensions. Moreover, comparison research indicates that DA-ViT surpasses state-of-the-art pre-trained Convolutional Neural Network (CNN) models in terms of accuracy and parameter efficiency.

---

Keywords

---