Hierarchical Data Integration With Gaussian Processes: Application to the Characterization of Cardiac...

Area of Science:

• Medical Imaging
• Machine Learning
• Cardiology

Background:

• Cardiac imaging generates diverse data, challenging statistical analysis.
• Existing fusion methods struggle with heterogeneous or correlated data.
• Physician expertise offers a hierarchical approach to data interpretation.

Purpose of the Study:

• To develop a novel hierarchical data fusion methodology for unsupervised representation learning in cardiac imaging.
• To mimic the physician's progressive integration of data based on established hierarchies.
• To enable more relevant statistical analysis of complex myocardial lesion patterns.

Main Methods:

• Proposed a Hierarchical Gaussian Process Latent Variable Model (GP-LVM).
• Integrated high-dimensional data descriptors progressively according to a known hierarchy.
• Linked latent representations and observations across hierarchical levels.

Main Results:

• Demonstrated the model's relevance on a dataset of 1726 cardiac MRI slices.
• Showcased consistent data organization across hierarchical levels.
• Achieved more relevant statistical analysis of myocardial lesion patterns, including microvascular obstruction (MVO).

Conclusions:

• The hierarchical fusion approach effectively organizes cardiac imaging data.
• The method enhances the statistical analysis of myocardial lesions, particularly subtle ones like MVO.
• This technique offers a promising framework for understanding cardiac conditions from imaging data.