Cortical surface registration using unsupervised learning.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Medical Image Analysis

Background:

• Non-rigid cortical registration is crucial for analyzing brain structure and function.
• Existing methods using spherical representations are computationally expensive.
• Direct application of deep learning to cortical surfaces is hindered by projection distortions.

Purpose of the Study:

• To develop an efficient and accurate deep learning framework for non-rigid cortical surface registration.
• To address the challenges of geometric complexity and inter-subject variability in cortical registration.
• To improve upon existing methods in terms of accuracy and computational cost.

Main Methods:

• Introduced SphereMorph, a diffeomorphic registration framework utilizing a UNet-style network with a spherical kernel.
• Employed a modified spatial transformer layer for sphere warping.
• Incorporated a resampling weight in the data fitting loss to mitigate polar projection distortions.

Main Results:

• SphereMorph demonstrates superior registration accuracy compared to conventional methods.
• The framework achieves significant improvements in computational efficiency.
• Successfully applied to cortical parcellation and group-wise functional area alignment tasks.

Conclusions:

• SphereMorph effectively models geometric registration problems within a convolutional neural network framework.
• The method offers a promising solution for fast and accurate cortical surface alignment.
• SphereMorph provides a valuable tool for advancing neuroimaging research.