Unsupervised multimodal surface registration with geometric deep learning

Area of Science:

• Neuroscience
• Computer Vision
• Medical Imaging

Background:

• Accurate image registration of cortical surfaces is crucial for understanding brain structure and function.
• Existing methods often struggle with complex surface geometries and achieving biologically plausible deformations.

Purpose of the Study:

• To introduce GeoMorph, a novel geometric deep-learning framework for robust and accurate cortical surface registration.
• To improve alignment quality and deformation smoothness in neuroimaging analysis.

Main Methods:

• Utilizes graph convolutions for independent feature extraction on cortical surfaces.
• Employs a deep-discrete registration approach with control point displacement learning.
• Incorporates a recurrent neural network-based conditional random field for regularization.

Main Results:

• GeoMorph demonstrates superior performance over existing deep-learning registration techniques.
• Achieves improved alignment accuracy and smoother, more biologically plausible deformations.
• Shows competitive results when compared to traditional registration frameworks.

Conclusions:

• GeoMorph offers a versatile and robust solution for cortical surface registration.
• Its advanced capabilities suggest significant potential for diverse neuroscience applications.
• The framework's code is publicly available for further research and development.