X-AstroNet: Explainable Adaptive Token Fusion via Hybrid Convolutional-Swin Transformer Pipelines for Alzheimer's Stage Identification

Abstract

The challenge of early detection of Alzheimer's Disease (AD) and Mild Cognitive Impairment (MCI) is a neuroimaging challenge that has not been solved yet, as conventional deep learning architectures are not well suited to extract fine-grained localized features from the image while preserving the long-range structural brain dependencies and maintaining interpretability. Here, we propose a novel 4-stage token fusion and self-attention solution for multi-class AD classification from structural Magnetic Resonance Imaging (sMRI) to overcome this challenge. The architecture combines a 3D Convolutional Neural Network (CNN) backbone for detecting subtle local structural abnormalities into a Swin Transformer for modelling global neuroanatomical relationships. The underlying principle of a Bio-inspired Astrocyte Gating Mechanism is the dynamic equilibrium between local and global representation tokens, and it is enhanced by an Explainable AI (XAI) component that produces high fidelity saliency maps for clinical validation. On benchmark MRI datasets, X-AstroNet achieves a classification accuracy of 91.79% and a sensitivity/recall of 91.79%, which is far better than the state-of-the-art baseline. This interpretable framework provides a highly reliable, clinically transparent tool for automated assessment of neurodegenerative diseases, by significantly reducing false-negative diagnoses across a range of early to severe diagnostic stages.