Brain graph synthesis by dual adversarial domain alignment and target graph prediction from a source graph

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Generating multimodal medical data from a single source is crucial for neurological disorder diagnosis, especially with limited data.
• Current deep learning models struggle with geometric data like brain graphs, hindering prediction from single-modality sources.
• Predicting target brain graphs from source brain graphs is an underexplored area, facing challenges like domain fracture.

Purpose of the Study:

• To develop a novel framework for predicting target brain graphs from source brain graphs.
• To address the domain fracture problem in multimodal medical data generation.
• To improve the accuracy and visual quality of synthesized brain graphs for enhanced diagnostic capabilities.

Main Methods:

• Proposed a Learning-guided Graph Dual Adversarial Domain Alignment (LG-DADA) framework.
• Implemented a source data pre-clustering step using manifold learning to manage data heterogeneity and prevent mode collapse.
• Utilized adversarial learning for domain alignment and dual adversarial regularization for joint embedding and prediction.

Main Results:

• The LG-DADA framework demonstrated superior prediction accuracy compared to existing graph synthesis methods.
• The method achieved better visual quality in the synthesized morphological brain graphs.
• Successfully addressed challenges related to domain fracture in graph prediction tasks.

Conclusions:

• The proposed LG-DADA framework effectively generates target brain graphs from source graphs.
• This approach offers a promising solution for improving neurological disorder diagnosis with reduced data acquisition.
• LG-DADA advances multimodal medical data synthesis for geometric data types.