Deep Computational Anatomy via Latent-Aligned Multiview Normalizing Flows

Area of Science:

• Computational anatomy
• Machine learning
• Medical imaging analysis

Background:

• Normalizing flows model complex probability distributions with exact likelihoods.
• Existing methods struggle with heterogeneous, multimodal datasets.
• Learning shared latent spaces is crucial for cross-modal data analysis.

Purpose of the Study:

• Introduce latent-aligned multiview normalizing (LAMNr) flows for multimodal data.
• Develop a framework for deep learning interpretations in computational anatomy.
• Enable exact cross-view imputation and latent space manipulation.

Main Methods:

• Utilize normalizing flows for bijective data mapping.
• Employ formal latent-alignment constraints to separate shared and view-specific features.
• Integrate with the ANTsX ecosystem (ANTsTorch) for PyTorch implementation.

Main Results:

• Demonstrated the framework's efficacy on imaging-derived phenotypes and multimodal MRI.
• Showcased potential for deep learning interpretations of computational anatomy concepts.
• Enabled closed-form conditional modeling for exact cross-view imputation.

Conclusions:

• LAMNr flows offer a powerful approach for multimodal data analysis.
• The framework provides a foundation for novel computational anatomy deep learning models.
• Open-source implementation facilitates further research and application in medical imaging.