Patho-radiological subtleties-aware multi-scale semantic fusion network for Alzheimer’s disease identification

Magnetic Resonance Imaging (MRI) enables the capture of subtle pathological differences closely linked to cognitive decline, providing a direct means of visualizing significant brain abnormalities in individuals with cognitive impairment. However, existing methods are insufficient for characterizing subtle pathological differences between individuals, such as regions of hippocampal atrophy or cortical thinning. Consequently, they fail to effectively model the complex functional and structural dependencies across brain regions during disease progression, making these subtle features difficult to detect and fully utilize. To tackle this problem, we proposed PSNet, a patho-radiological subtleties aware multi-scale semantic fusion network for Alzheimer’s Disease (AD) identification. Firstly, to accurately capture subtle pathological differences across brain regions at multiple scales, we proposed the pathological local subtle differences feature extraction (PLSDF), which combined the Lightweight Pyramid Squeeze Attention (LPSA) mechanism with a residual network. This design enabled the perception of subtle differences in local pathological features and the acquisition of cross-hierarchical information on lesion distribution across brain regions. Secondly, to enhance global dependencies among the brain’s pathological features, we proposed multi-scale semantics feature extraction of MRI based on modified ViT (MSFE). This approach employed a feature reuse strategy that used atrous convolutions to enhance subtle local pathological features, pooling operations to capture global contextual information about the brain’s overall structure, and the multi-head attention mechanism to model pathological dependencies across different brain regions associated with AD. Finally, the experiments demonstrated that PSNet achieves superior performance compared to representative methods on three publicly available datasets. Code is available at: https://github.com/xub-code/PSNet.git .