Parameter Space CNN for Cortical Surface Segmentation.

Area of Science:

• Neuroimaging
• Deep Learning
• Computational Neuroscience

Background:

• Spherical coordinate systems are standard for analyzing human cortical neuroimaging data.
• Surface-based signals like curvature and myelination define cortical regions.
• Current surface-based deep learning (spherical CNNs) struggles with accurate cortical segmentation.

Purpose of the Study:

• To introduce and evaluate a novel 2D parameter space approach with view aggregation (p3CNN) for human cortical segmentation.
• To improve the accuracy of surface-based deep learning for cortical segmentation tasks.

Main Methods:

• Development and evaluation of the p3CNN network, a 2D parameter space approach with view aggregation.
• Comparison of p3CNN performance against spherical CNNs for cortical segmentation.

Main Results:

• The p3CNN network significantly outperforms spherical CNNs in cortical segmentation accuracy.
• The average Dice similarity score for cortical segmentation using p3CNN exceeded 0.9.

Conclusions:

• The p3CNN approach offers a more accurate method for surface-based segmentation of human cortical neuroimaging data.
• This advancement addresses limitations in current deep learning models for neuroimaging analysis.