BrainUMA: A Unified multi-atlas learning framework for brain disorders diagnosis.

Area of Science:

• Neuroscience
• Medical Imaging
• Machine Learning

Background:

• Functional connectivity analysis aids brain disorder diagnosis.
• Multi-atlas approaches enhance diagnostic accuracy but face challenges with atlas interaction and data redundancy.
• Existing methods often lack sufficient interaction and consistency among multiple brain atlases.

Purpose of the Study:

• To propose a unified multi-atlas learning framework (BrainUMA) for improved brain disorder diagnosis.
• To enhance the modeling of interactions and consistency across multiple brain atlases.
• To investigate the optimal combination of brain atlases for diagnostic tasks.

Main Methods:

• Developed a unified multi-atlas learning framework (BrainUMA) incorporating hyper-connectivity network learning.
• Implemented a novel hyper-connectivity network construction strategy involving hypergraph structure and node feature learning.
• Utilized feature disentanglement with hyperedge-aware hypergraph convolutional networks and introduced contrastive and class-consistency losses.

Main Results:

• Demonstrated the effectiveness of BrainUMA on the Autism Brain Imaging Data Exchange (ABIDE) dataset.
• Showcased the importance of disentanglement for improving multi-atlas disease diagnosis.
• Provided deeper insights into disease interpretability, including atlas properties and critical brain regions.

Conclusions:

• BrainUMA offers a robust framework for brain disorder diagnosis by effectively leveraging multi-atlas information.
• Feature disentanglement is crucial for maximizing the benefits of multi-atlas learning in neuroimaging.
• The model enhances diagnostic accuracy and interpretability, offering valuable insights into brain network alterations in disease.