HDFT-MViT: a progressive core-enhanced mix framework for Alzheimer's disease classification using MRI images

Early and accurate diagnosis of Alzheimer's disease (AD) is critical. In MRI-based computer-aided diagnosis, convolutional neural networks (CNNs) excel at extracting local features but struggle to model long-range dependencies, while Vision Transformers (ViTs) offer strong global modeling capabilities but suffer from high computational complexity, limiting their deployment in resource-constrained settings.

This paper proposes HDFT-MViT, a lightweight hybrid architecture based on MobileViT that integrates a hierarchical dynamic filter with a lightweight Transformer. The model adopts a progressive Core-Enhanced Mix design: Shallow layers employ MobileNetV2 inverted residual blocks for efficient local feature extraction; intermediate and deep layers incorporate a dual-branch module that integrates a dynamic filter for frequency-domain global modulation and a lightweight Transformer for spatial long-range dependency modeling, followed by hierarchical fusion via learnable weights. A channel attention mechanism is further introduced to enhance feature discriminability.

Evaluations on the public ADNI-1 (3-class) and ADNI-2 (4-class) MRI datasets show that HDFT-MViT achieves state-of-the-art classification accuracies of 98.85 ± 0.27% and 98.07 ± 0.54%, respectively, while maintaining a lightweight profile with only 3.46 M parameters, confirming its effectiveness and efficiency.

HDFT-MViT achieves an optimal balance between local detail perception and global semantic understanding within a computationally efficient framework, offering a promising tool for clinical AD diagnosis. Code will be released upon acceptance.