Motivation: Accurate analysis of the cerebral cortex's complex geometry is crucial for studying various brain disorders. However, its intricate folds make segmentation challenging and prone to topological errors. Goal(s): We proposed a semi-supervised framework to effectively correct topological errors in the cerebral cortex. Approach: Our framework uses a generation network to estimate pseudo ground truth, enabling semi-supervised training of a topological correction network. During testing, only the trained correction network is required to directly produce correction results. Results: We evaluated the proposed framework on 165 lifespan images and demonstrated that it significantly outperforms several state-of-the-art correction methods. Impact: The proposed semi-supervised framework addresses topological defects in the segmentation of the brain's complex folds, providing improved accuracy in cortical analysis and surpassing existing correction methods. This advancement has the potential to enhance studies of neurodevelopmental, neurodegenerative, and psychological disorders.
Sun et al. (2025) studied this question.