Dynamic Hyper-Graph Inference Framework for Computer-Assisted Diagnosis of Neurodegenerative Diseases

Area of Science:

• Medical Imaging
• Computer Vision
• Graph Theory

Background:

• Hyper-graph methods excel in multi-modal neuroimaging but struggle with data consistency and cross-modality integration.
• Existing approaches generate representations independently of label inference, leading to suboptimal predictions.
• Sequential hyper-graph construction and voting limit the exploitation of complementary multi-modal data relationships.

Purpose of the Study:

• To propose a novel dynamic hyper-graph inference method using a semi-supervised framework.
• To enhance the consistency between learned hyper-graph structures and observed clinical labels/scores.
• To integrate classification and regression tasks within a unified framework for neuroimaging analysis.

Main Methods:

• Developed a dynamic hyper-graph inference method with a semi-supervised framework.
• Iteratively estimated and adjusted hyper-graph structures for multi-modal data.
• Integrated classification and regression for disease identification and score prediction.

Main Results:

• Achieved improved performance in identifying mild cognitive impairment (MCI) subjects.
• Demonstrated enhanced fine-grained recognition of MCI progression stages.
• Outperformed conventional hyper-graph inference methods in experimental evaluations.

Conclusions:

• The proposed dynamic hyper-graph inference method offers superior performance for multi-modal neuroimaging data.
• The framework effectively aligns hyper-graph structures with clinical phenotypes.
• This approach provides a unified and robust solution for neuroimaging-based disease classification and regression.