Cortical Surface Parcellation using Spherical Convolutional Neural Networks.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Artificial Intelligence

Background:

• Traditional multi-atlas cortical surface parcellation is time-consuming (2-3 hours per subject).
• Existing methods struggle with optimal parcel boundary matching to geometric features.
• Accurate cortical parcellation relies heavily on the selection of appropriate training features.

Purpose of the Study:

• To develop an efficient and accurate cortical surface parcellation method.
• To overcome the limitations of traditional registration-based parcellation techniques.
• To leverage deep learning for improved brain mapping.

Main Methods:

• Utilized spherical deep convolutional neural networks for cortical parcellation.
• Proposed novel parcellation-specific input data tailored for complex cortical surface structures.
• Generated new training data by aligning ground-truth parcel boundaries and using deformation fields to create morphed features and maps.

Main Results:

• The proposed method significantly outperforms traditional multi-atlas and naive spherical U-Net approaches.
• Achieved full cortical parcellation in less than one minute per subject.
• Validated on 427 adult brains for 49 distinct cortical labels.

Conclusions:

• Spherical deep convolutional neural networks offer a highly efficient and accurate solution for cortical surface parcellation.
• The novel data augmentation strategy enhances network performance for complex neuroanatomical data.
• This method represents a substantial advancement in automated brain mapping and analysis.