UNSUPERVISED CORTICAL SURFACE REGISTRATION NETWORK FOR ALIGNING GYRALNET.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Medical Image Analysis

Background:

• The 3-hinge gyrus (3HG) and its network (GyralNet) are crucial for understanding brain structure and function.
• Current cortical surface registration methods do not adequately account for 3HG and GyralNet, leading to suboptimal alignment.
• Existing atlases for 3HG and GyralNet are unavailable, and traditional atlas generation is time-consuming.

Purpose of the Study:

• To develop an unsupervised learning framework for jointly creating 3HG and GyralNet atlases.
• To enable accurate registration of individual cortical features onto the generated atlas.
• To improve the alignment of brain structures by incorporating 3HG and GyralNet information.

Main Methods:

• An unsupervised learning framework was developed to jointly generate 3HG and GyralNet atlases.
• 3HG and GyralNet graph structures were converted into surface distance maps for integration into the registration network.
• A multi-level spherical registration network, based on spherical U-Net, was employed for coarse-to-fine registration to handle large deformations.

Main Results:

• The proposed framework successfully generated detailed 3HG and GyralNet atlases.
• The method demonstrated improved registration accuracy compared to existing approaches.
• The integration of surface distance maps effectively incorporated the graph structure of 3HG and GyralNet.

Conclusions:

• The unsupervised learning framework provides an efficient and effective method for generating 3HG and GyralNet atlases.
• This approach significantly enhances the accuracy of cortical surface registration by accounting for key gyral features.
• The developed framework offers a valuable tool for neuroimaging research requiring precise brain structure alignment.