SrvfNet: A Generative Network for Unsupervised Multiple Diffeomorphic Functional Alignment.

Area of Science:

• Computational neuroscience
• Medical imaging analysis
• Machine learning

Background:

• Functional data analysis requires robust alignment methods.
• Current methods may lack unsupervised capabilities or template prediction.
• Square-root velocity functions (SRVF) offer a powerful representation for functional data.

Purpose of the Study:

• Introduce SrvfNet, a novel generative deep learning framework.
• Enable unsupervised joint multiple alignment of functional data using SRVF.
• Facilitate alignment to predefined templates and simultaneous optimal template prediction.

Main Methods:

• Developed a generative encoder-decoder deep learning architecture.
• Utilized fully-connected layers to model warping function distributions.
• Applied the framework to large collections of SRVF data.

Main Results:

• Successfully aligned synthetic functional data.
• Demonstrated effective alignment on magnetic resonance imaging (MRI) diffusion profiles.
• Validated the unsupervised and template prediction capabilities.

Conclusions:

• SrvfNet provides an effective unsupervised deep learning solution for SRVF alignment.
• The framework shows promise for analyzing complex functional datasets, including medical imaging.
• Future work can explore extensions to other functional data types and applications.