Spherical deformation method with SO(3)-equivariant model boosts sulcal labeling of small cortical regions

An isotropic spherical region enlargement method using SO(3)-equivariant deep learning improved Dice score for the smallest sulcus by 11.16% over baseline, with performance gains negatively correlated with ROI size, confirming its utility for labeling small cortical sulci.

• The method uses an exponential expansion function on spherical surfaces to enlarge sulcal ROIs and contract unlabeled regions, with an SO(3)-equivariant model to improve generalization.
• In experiments on adult and pediatric cohorts, the method outperformed baselines, with the largest improvement of 11.16% Dice score for a hominoid-specific sulcus implicated in cognition and disorders.
• Performance gains were strongly negatively correlated with ROI size, indicating particular effectiveness for small sulci.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.