Constructing hierarchical attentive functional brain networks for early AD diagnosis

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Diagnostics

Background:

• Conventional functional brain network (FBN) analysis relies on single-scale representations and predefined brain atlases.
• The brain's hierarchical organization necessitates multiscale analysis for comprehensive understanding.
• Existing methods lack the flexibility to adapt to the brain's inherent hierarchical structure.

Purpose of the Study:

• To develop an adaptive hierarchical FBN construction method within the Transformer framework.
• To integrate complementary diagnostic information across multiple scales of brain activity.
• To overcome the limitations of predefined brain atlases by exploring finer-grained network representations.

Main Methods:

• Proposed a Transformer-based framework for adaptive hierarchical FBN construction.
• Introduced a sparse attention-based node-merging module to generate coarser network nodes.
• Implemented a node-splitting strategy to explore finer-grained representations beyond traditional atlases.

Main Results:

• The proposed method adaptively learns hierarchical FBN representations in an end-to-end manner.
• Hierarchical structure effectively integrates multiscale FBN information, reducing model complexity.
• Experiments demonstrated superior performance compared to existing methods across various datasets.

Conclusions:

• Adaptive hierarchical FBN construction offers a more effective approach for brain disorder diagnosis.
• The Transformer-based method captures complementary insights from different network scales.
• Exploring finer-grained nodes and adaptive hierarchy enhances FBN representation and diagnostic utility.