Bidirectional Mapping with Contrastive Learning on Multimodal Neuroimaging Data.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Deep learning models show promise in identifying biomarkers for brain diseases by analyzing brain structure and function.
• Current methods often use unidirectional mapping, which can introduce bias and overlook the integrated nature of brain structure and function.

Purpose of the Study:

• To develop a novel bidirectional mapping model to address the limitations of unidirectional approaches.
• To reduce bias in mapping brain structure to function and vice versa.
• To improve the prediction of clinical phenotypes and neurodegenerative diseases.

Main Methods:

• Proposed a novel Bidirectional Mapping with Contrastive Learning (BMCL) model.
• Employed ROI-level contrastive learning to reduce bias between unidirectional mappings.
• Evaluated the framework on clinical phenotype and neurodegenerative disease prediction tasks.

Main Results:

• The BMCL model demonstrated superior performance compared to existing state-of-the-art methods.
• The bidirectional approach effectively reduced bias inherent in unidirectional mapping.
• Accurate predictions for clinical phenotypes and neurodegenerative diseases were achieved.

Conclusions:

• The BMCL model offers a more robust and less biased approach to modeling brain structure-function interactions.
• This framework has significant potential for advancing biomarker discovery and disease prediction in neuroscience.
• Bidirectional mapping is crucial for accurately representing the complex relationship between brain structure and function.