From voxel discovery to regional interaction: A multi-level interpretable framework for Alzheimer's disease diagnosis

Alzheimer's disease (AD) is an irreversible neurodegenerative disorder where early diagnosis serves as the only viable window for effective intervention. While Structural MRI (sMRI) is a primary clinical tool for this purpose, manual diagnosis is heavily constrained by clinician expertise and the difficulty of visually quantifying subtle, diffuse anatomical changes. Deep learning has emerged as a powerful solution to capture these complex patterns. However, the inherent black-box opacity of current models creates a trust deficit that severely impedes their clinical adoption. Furthermore, existing interpretability methods predominantly rely on post-hoc visualizations, which frequently suffer from significant background leakage and lack anatomical coherence. To overcome these barriers, we propose a novel multi-level interpretable framework that inherently integrates diagnostic reasoning into the model architecture. The framework operates through a progressive three-stage trajectory to bridge the gap between pixel-level signals and global pathology. First, to eliminate the ambiguity caused by background leakage, we introduce a voxel discovery module implemented via Hierarchical Evidential Masking (HiEM). By leveraging this iterative optimization strategy, the model actively filters out irrelevant background signals to autonomously discover AD-relevant voxel sets. Subsequently, to bridge the semantic gap between discrete voxels and clinical anatomy, these discovered signals are dynamically aggregated into coherent semantic regions. Finally, recognizing that AD pathology propagates through brain networks, we explicitly model the regional interaction among these defined areas using Graph Attention Network (GAT). This voxel-region-network progression allows the model to transparently zoom out from local cues to global interactions. Validated on the Alzheimer's Disease Neuroimaging Initiative (ADNI) and Australian Imaging, Biomarkers and Lifestyle Study (AIBL) datasets, our method achieves performance comparable to state-of-the-art black-box models while autonomously reconstructing established neuropathological trajectories. Crucially, by extracting high-fidelity pathological signals, it offers a trustworthy tool for precise disease diagnosis and potentially serves as a digital biomarker discovery instrument for localizing subtle prodromal lesions. The source code is publicly available at: https://github.com/k-Jayus/HiEM.