Multi-Atlas Brain Network Classification Through Consistency Distillation and Complementary Information Fusion.

Area of Science:

• Neuroscience
• Medical Imaging
• Computational Biology

Background:

• Brain network analysis using functional magnetic resonance imaging (fMRI) is vital for identifying neurological disorder patterns.
• Current methods struggle with inconsistencies and lack of information exchange across multiple brain atlases.
• Existing approaches limit the detection of abnormalities due to the absence of a standard brain atlas.

Purpose of the Study:

• To propose a novel framework, AIDFusion, for enhanced brain network classification using fMRI data.
• To address limitations in multi-atlas consistency and ROI-level information exchange in current methods.
• To improve the accuracy and interpretability of brain network analysis for neurological disorder identification.

Main Methods:

• Developed the Atlas-Integrated Distillation and Fusion network (AIDFusion) incorporating a disentangle Transformer.
• Implemented cross-atlas information filtering, distillation, and fusion mechanisms.
• Applied subject- and population-level consistency constraints for improved cross-atlas coherence.

Main Results:

• AIDFusion demonstrated superior classification performance across four resting-state fMRI datasets.
• The framework showed significant computational efficiency compared to existing state-of-the-art methods.
• Case studies revealed AIDFusion's ability to extract interpretable patterns consistent with neuroscience findings.

Conclusions:

• AIDFusion offers a robust solution for multi-atlas brain network analysis in the context of neurological disorders.
• The framework enhances the accuracy and interpretability of fMRI-based brain network classification.
• AIDFusion represents a significant advancement in leveraging multi-atlas information for clinical neuroscience research.