TASNet: A tri-modal attentive scale robust adaptive fusion framework for progressive mild cognitive impairment classification.
Authors
Affiliations (2)
Affiliations (2)
- School of Computer and Artificial Intelligence, Shandong Jianzhu University, Jinan, Shandong, China.
- School of Computer and Artificial Intelligence, Shandong Jianzhu University, Jinan, Shandong, China. Electronic address: [email protected].
Abstract
Accurate prediction of progressive Mild Cognitive Impairment (pMCI) versus stable MCI (sMCI) is crucial for early Alzheimer's disease (AD) intervention and delaying disease progression. The combination of Positron Emission Tomography (PET) and structural Magnetic Resonance Imaging (sMRI) is currently the most recognized diagnostic basis. However, PET and sMRI reveal lesions with heterogeneous scales, irregular spatial distributions, and inter-modal temporal asynchrony-factors causing inadequate feature extraction and poor fusion in existing methods. To address these challenges, this paper introduces Tri-modal Attention Scale-robust Network (TASNet), an innovative feature-scale robust framework integrating brain imaging (PET, sMRI) with Cerebrospinal Fluid (CSF) data. TASNet adaptively fuses these modalities via symmetric dynamic channel attention in an end-to-end manner. A 3D Parallel Axial Fusion Attention (PAFA) mechanism and a Distribution-Variable Atrous Spatial Pyramid Pooling (DV-ASPP) module are designed to efficiently capture multi-scale irregular pathological features. Additionally, a Hardness-Aware Hybrid Loss (HAHL) function is introduced to enhance the discriminability of challenging cases. Experiments on the ADNI dataset resulted in an accuracy of 85.00% and an AUC of 91.67% in pMCI/sMCI classification. Furthermore, explainability analysis confirms the model focuses on AD pathology-consistent regions (e.g., temporal and parietal lobes). The proposed TASNet demonstrates superior performance compared to state-of-the-art methods. By mitigating information loss typical of conventional serial models and effectively handling multi-scale features, the framework achieves higher classification metrics than existing approaches. This work provides a technical solution for early AD prediction with high accuracy, reliability, and clinical interpretability.