BrainDAS: Structure-aware domain adaptation network for multi-site brain network analysis.

Area of Science:

• Neuroscience
• Medical Informatics
• Machine Learning

Background:

• Multi-site medical datasets are common due to data acquisition challenges, but lead to domain shift issues.
• Heterogeneous data distributions across sites hinder accurate identification of conditions like autism spectrum disorder (ASD).
• Domain adaptation is a promising solution, but its application to graph data, like brain networks, remains understudied.

Purpose of the Study:

• To propose an end-to-end structure-aware domain adaptation framework, BrainDAS, for analyzing brain networks from resting-state functional magnetic resonance imaging (rs-fMRI).
• To address the challenges of complex graph structures and multiple source domains in graph-based domain adaptation.
• To improve the identification of autism spectrum disorder (ASD) in multi-site datasets.

Main Methods:

• Developed BrainDAS, a two-stage framework incorporating supervision-guided multi-site graph domain adaptation with dynamic kernel generation.
• Implemented attention-based graph pooling for graph classification within the framework.
• Utilized the Autism Brain Imaging Data Exchange (ABIDE) dataset, comprising 871 subjects from 17 sites, for evaluation.

Main Results:

• BrainDAS outperformed state-of-the-art algorithms in various evaluation settings on the ABIDE dataset.
• The framework demonstrated promising interpretability and generalization capabilities.
• Achieved significant improvements in identifying autism spectrum disorder across diverse data sources.

Conclusions:

• The proposed BrainDAS framework effectively handles domain shift in multi-site rs-fMRI data for brain network analysis.
• BrainDAS offers a robust and interpretable solution for ASD identification, outperforming existing methods.
• The framework's success highlights the potential of structure-aware domain adaptation for complex graph data in medical imaging.