LGL‐Net: A Lightweight Global‐Local Multiscale Network With Region‐Aware Interpretability for Alzheimer's Disease Diagnosis

ABSTRACT Alzheimer's disease (AD) is a progressive neurodegenerative disorder marked by gradual cognitive decline and structural brain degeneration. Magnetic resonance imaging (MRI), due to its non‐invasive nature and high spatial resolution, plays a pivotal role in the clinical diagnosis of AD. However, considerable challenges persist, primarily due to the heterogeneity of brain structural alterations across individuals and the high computational burden associated with deploying deep learning models in clinical practice. Although recent deep learning‐based approaches have significantly improved diagnostic accuracy, most models fail to identify the specific contributions of individual brain regions, limiting their interpretability and clinical applicability. To address these limitations, we propose LGL‐Net, a novel lightweight 3D convolutional neural network tailored for efficient extraction and integration of both global and local anatomical features from MRI data. The architecture adopts a dual‐branch design, wherein one branch captures whole‐brain atrophy patterns, while the other focuses on fine‐grained, region‐specific structural variations. This design achieves a favourable trade‐off between computational efficiency and diagnostic performance, significantly reducing the model's parameter count and computational load without compromising accuracy. Importantly, LGL‐Net explicitly maps learnt features onto anatomically defined brain regions, enabling region‐level interpretability of classification outcomes. By independently evaluating the contributions of each region to both global and local representations, the model elucidates how multiscale anatomical features collectively influence diagnostic decisions. Experimental results demonstrate that LGL‐Net achieves classification performance comparable to existing methods, while substantially lowering model complexity and computational demands. Overall, this framework offers a scalable, interpretable and resource‐efficient solution for intelligent AD diagnosis.