SegCSR: WEAKLY-SUPERVISED CORTICAL SURFACES RECONSTRUCTION FROM BRAIN RIBBON SEGMENTATIONS.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Computational Neuroscience

Background:

• Deep learning cortical surface reconstruction (CSR) relies on pseudo ground truth (pGT), causing dataset-specific issues and extensive data preparation.
• Existing supervised methods require meticulously prepared training data, limiting scalability and generalizability.

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 pGT-dependent supervised learning in CSR.

Main Methods:

• SegCSR initializes a midthickness surface and deforms it to white matter and pial surfaces using learned diffeomorphic flows.
• Joint learning incorporates boundary surface loss for alignment and inter-surface normal consistency loss for regularization in deep sulci.
• Additional terms enforce surface smoothness and correct topology.

Main Results:

• Weakly-supervised SegCSR achieves comparable or superior accuracy and regularity to supervised deep learning CSR methods.
• The method was validated on two large-scale brain MRI datasets.
• Demonstrates effective reconstruction of inner and outer cortical surfaces.

Conclusions:

• SegCSR offers an effective weakly-supervised alternative for cortical surface reconstruction.
• Reduces the need for pGT, simplifying data preparation and potentially improving generalizability.
• Presents a promising direction for advancing automated brain MRI analysis.