From Symptomatic to Pre-Symptomatic: Adaptive Knowledge Distillation for Early Alzheimer's Detection Using Functional...

Area of Science:

• Neuroscience
• Medical Imaging
• Artificial Intelligence

Background:

• Alzheimer's disease (AD) diagnosis requires early detection for timely intervention.
• Functional magnetic resonance imaging (fMRI) shows promise for non-invasive AD biomarkers.
• Current fMRI methods struggle with pre-symptomatic AD detection due to subtle patterns and severe class imbalance.

Purpose of the Study:

• To develop a novel framework for reliable pre-symptomatic Alzheimer's disease detection using fMRI.
• To address challenges of subtle anatomical differences and class imbalance in early AD detection.
• To improve diagnostic accuracy by leveraging knowledge distillation from symptomatic to pre-symptomatic stages.

Main Methods:

• Proposed a margin-aware knowledge distillation (KD) framework for multi-stage AD diagnosis.
• Utilized Unbalanced Optimal Transport (UOT) for feature distillation to adapt to neurodegeneration-induced anatomical changes.
• Implemented Self-Distillation with Dynamic Margins to mitigate class imbalance issues.

Main Results:

• Demonstrated the superiority of the proposed KD framework over state-of-the-art methods across four base models.
• Successfully identified significant brain regions involved in pre-symptomatic AD detection.
• Showcased effective feature transfer during the distillation process.

Conclusions:

• The novel KD framework significantly advances early Alzheimer's disease diagnosis capabilities.
• The approach offers more precise diagnostic tools by understanding early disease manifestations.
• This work represents a significant stride towards reliable and earlier Alzheimer's disease detection.