Spherical Deformable U-Net: Application to Cortical Surface Parcellation and Development Prediction.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Computational Geometry

Background:

• Convolutional Neural Networks (CNNs) excel in Euclidean image analysis but struggle with spherical topology common in medical imaging, like brain cortical surfaces.
• Standard CNN operations lack consistent neighborhood definitions for mesh-based cortical surface data, hindering analysis.
• Existing methods for spherical data analysis are limited in modeling geometric transformations.

Purpose of the Study:

• To develop novel convolution and pooling operations for spherical cortical surface data.
• To introduce a Spherical Deformable U-Net (SDU-Net) for enhanced geometric transformation modeling on spherical data.
• To apply SDU-Net to neuroimaging tasks for improved accuracy and efficiency.

Main Methods:

• Leveraged regular, hierarchical geometric structure of resampled spherical cortical surfaces.
• Developed 1-ring filter, convolution, and pooling operations for Spherical U-Net.
• Introduced deformable convolution and pooling with learned spherical offsets for adaptive localization, creating SDU-Net.

Main Results:

• SDU-Net adaptively localizes cortical structures of varying sizes and shapes on the sphere.
• Achieved competitive performance in accuracy and computational efficiency for cortical surface parcellation and attribute map prediction.
• Demonstrated the effectiveness of learned spherical offsets in enhancing transformation modeling.

Conclusions:

• SDU-Net offers a powerful new approach for analyzing complex spherical data in medical imaging.
• The method significantly advances neuroimaging by improving the analysis of brain cortical surfaces.
• SDU-Net provides a more robust and efficient tool for tasks like brain parcellation and mapping cortical attributes.