TASNet: A tri-modal attentive scale robust adaptive fusion framework for progressive mild cognitive impairment...

Area of Science:

• Neuroimaging
• Biomarker Analysis
• Artificial Intelligence in Medicine

Background:

• Accurate prediction of progressive Mild Cognitive Impairment (pMCI) versus stable MCI (sMCI) is critical for early Alzheimer's disease (AD) intervention.
• Current methods using PET and sMRI face challenges due to heterogeneous scales, irregular distributions, and temporal asynchrony of lesions, leading to inadequate feature extraction and fusion.
• Integrating Cerebrospinal Fluid (CSF) data with neuroimaging offers a potential avenue for improved diagnostic accuracy.

Purpose of the Study:

• To develop an innovative, feature-scale robust framework for accurate pMCI vs. sMCI classification.
• To address limitations in existing methods for extracting and fusing multi-modal imaging data (PET, sMRI) and CSF biomarkers.
• To enhance the discriminability of challenging cases in early AD prediction.

Main Methods:

• Introduction of the Tri-modal Attention Scale-robust Network (TASNet), an end-to-end framework integrating PET, sMRI, and CSF data.
• Utilizing a 3D Parallel Axial Fusion Attention (PAFA) mechanism and Distribution-Variable Atrous Spatial Pyramid Pooling (DV-ASPP) for multi-scale pathological feature capture.
• Employing a Hardness-Aware Hybrid Loss (HAHL) function to improve classification of difficult cases and adaptive fusion via symmetric dynamic channel attention.

Main Results:

• TASNet achieved 85.00% accuracy and 91.67% AUC on the ADNI dataset for pMCI/sMCI classification.
• Explainability analysis confirmed the model's focus on brain regions consistent with AD pathology, such as the temporal and parietal lobes.
• The proposed method demonstrated superior performance compared to state-of-the-art approaches, mitigating information loss and handling multi-scale features effectively.

Conclusions:

• TASNet offers a robust technical solution for early AD prediction with high accuracy and reliability.
• The framework provides clinical interpretability by focusing on pathology-relevant brain regions.
• This approach enhances the classification metrics for pMCI/sMCI prediction, outperforming existing methods.