APLG-Net: an anatomy-guided local-global hybrid network with progression-aware supervision for structural MRI-based...

Area of Science:

• Neuroimaging
• Artificial Intelligence in Medicine
• Biomedical Engineering

Background:

• Structural MRI-based classification of Alzheimer's disease (AD) is challenging due to subtle anatomical differences and the difficulty in distinguishing mild cognitive impairment (MCI).
• Accurate classification is crucial for timely intervention and disease management.

Purpose of the Study:

• To develop an advanced deep learning model, APLG-Net, for improved classification of Normal Control (NC), MCI, and AD using structural MRI.
• To enhance the model's ability to capture both global and local brain structural information and incorporate disease progression patterns.

Main Methods:

• Proposed APLG-Net, a hybrid network combining a global whole-brain encoder and a local region-of-interest (ROI)-based encoder.
• Implemented cross-attention fusion and vector-gated integration for feature combination.
• Introduced an ordinal supervision strategy to model disease progression for enhanced classification.

Main Results:

• APLG-Net achieved 87.1% accuracy, 86.4% balanced accuracy, and 86.8% Macro-F1 on the ADNI dataset.
• Specifically, it reached an 85.6% F1 score for MCI classification.
• Outperformed existing CNN-based, Transformer-based, and other hybrid baseline models.

Conclusions:

• The integration of anatomical priors, local-global feature interaction, and ordinal supervision significantly boosts classification performance.
• APLG-Net demonstrates superior robustness and improved discrimination for MCI, a critical stage in Alzheimer's disease progression.