SegCSR: Weakly-Supervised Cortical Surfaces Reconstruction from Brain Ribbon Segmentations.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Computational Neuroscience

Background:

• Deep learning for cortical surface reconstruction (CSR) typically requires pseudo ground truth (pGT) data.
• This reliance on pGT leads to dataset-specific issues and extensive data preparation.

Purpose of the Study:

• To develop a weakly-supervised deep learning method for reconstructing multiple cortical surfaces from brain MRI.
• To overcome the limitations of traditional supervised CSR methods.

Main Methods:

• Initialize a midthickness surface and deform it to inner (white matter) and outer (pial) surfaces using learned diffeomorphic flows.
• Employ a boundary surface loss to align surfaces with segmentation map boundaries.
• Utilize inter-surface normal consistency loss for regularization in deep sulci.
• Incorporate regularization for surface smoothness and topology.

Main Results:

• The weakly-supervised method achieves CSR accuracy comparable or superior to supervised deep learning alternatives.
• Evaluated on two large-scale brain MRI datasets, demonstrating robust performance.
• The method shows improved regularity in cortical surface reconstruction.

Conclusions:

• Weakly-supervised learning offers a viable and effective alternative for cortical surface reconstruction.
• This approach reduces reliance on pGT, simplifying data preparation and enhancing generalizability.
• The proposed method provides accurate and regular cortical surface reconstructions for neuroimaging research.