SPHARM-Net: Spherical Harmonics-Based Convolution for Cortical Parcellation.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Machine Learning

Background:

• Cortical parcellation is crucial for understanding brain structure and function.
• Existing Convolutional Neural Networks (CNNs) for parcellation often rely on spatial convolutions, necessitating extensive data augmentation and limiting neighborhood definitions.
• These limitations can hinder accuracy and efficiency in brain mapping.

Purpose of the Study:

• To develop a novel spherical harmonics-based CNN, SPHARM-Net, for accurate and efficient cortical parcellation.
• To address limitations of existing methods by proposing a constrained spherical convolutional filter and an end-to-end framework.
• To achieve rotational equivariance and reduce reliance on data augmentation.

Main Methods:

• A constrained spherical convolutional filter supporting an infinite set of spectral components, encoding all components without full spherical harmonics expansion.
• An end-to-end framework for cortical parcellation that does not require data augmentation.
• Utilizing matrix transformations for efficient and fast spectral processing, enabling rotational equivariance.

Main Results:

• SPHARM-Net demonstrated significantly reduced training times due to rotational equivariance.
• The method achieved accurate cortical parcellation on the Mindboggle-101 and NAMIC datasets.
• SPHARM-Net outperformed state-of-the-art methods with fewer learnable parameters and without rigid alignment or data augmentation.

Conclusions:

• SPHARM-Net offers a computationally efficient and accurate approach to cortical parcellation.
• The proposed constrained spherical convolutional filter and framework overcome key limitations of previous CNN-based methods.
• The method shows promise for advancing neuroimaging analysis and brain mapping.