CortexODE: Learning Cortical Surface Reconstruction by Neural ODEs.

Area of Science:

• Computational Neuroscience
• Medical Image Analysis
• Machine Learning

Background:

• Cortical surface reconstruction is crucial for understanding brain structure and function.
• Existing methods are often slow and computationally intensive, limiting their clinical application.
• Deep learning offers potential for accelerating and improving reconstruction accuracy.

Purpose of the Study:

• To introduce CortexODE, a novel deep learning framework for efficient and accurate cortical surface reconstruction.
• To leverage neural ordinary differential equations (ODEs) for learning diffeomorphic surface flows.
• To develop an automated pipeline for reconstructing white matter and pial surfaces from brain MRI.

Main Methods:

• Utilized a 3D U-Net for white matter segmentation and initial surface generation via signed distance functions.
• Employed neural ODEs with a Lipschitz-continuous deformation network to model surface point trajectories.
• Integrated fast topology correction and isosurface extraction for robust surface generation.

Main Results:

• The CortexODE pipeline reconstructs cortical surfaces in under 5 seconds.
• Achieved an average geometric error of less than 0.2mm across diverse age groups.
• Demonstrated theoretical guarantees against surface self-intersections due to the ODE formulation.

Conclusions:

• CortexODE provides a significant advancement in the speed and accuracy of cortical surface reconstruction.
• The framework is robust, validated on large-scale datasets spanning neonates to the elderly.
• This deep learning approach has the potential to revolutionize neuroimaging analysis pipelines.