Cycle-consistent Learning for Fetal Cortical Surface Reconstruction.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Developmental Neuroscience

Background:

• Accurate fetal cortical surface reconstruction is vital for analyzing prenatal brain development.
• Existing methods are insufficient for fetal brain MRI due to unique challenges like low contrast, motion artifacts, and rapid folding.
• Fetal brains exhibit narrow cortical ribbons and sulci, increasing susceptibility to partial volume effects and boundary ambiguities.

Purpose of the Study:

• To develop a novel deep learning-based method for fetal cortical surface reconstruction.
• To address the scarcity of specialized techniques for fetal brain imaging.
• To improve the accuracy and reliability of quantitative analysis in prenatal neurodevelopment.

Main Methods:

• A multi-task, prior-knowledge supervised deep learning framework was developed.
• Incorporated a cycle-consistent strategy with prior knowledge and stationary velocity fields.
• Employed iterative refinement of inner and outer cortical surfaces through mutual guidance for enhanced accuracy.

Main Results:

• Achieved a geometric error of 0.229 ± 0.047 mm on an 83-subject fetal MRI dataset.
• Demonstrated a low self-intersecting face rate of 0.023 ± 0.058%, indicating high topological accuracy.
• Outperformed state-of-the-art deep learning methods in accuracy and computational efficiency.

Conclusions:

• The developed deep learning method significantly advances fetal cortical surface reconstruction.
• The technique provides high geometric and topological accuracy, crucial for prenatal neurodevelopmental studies.
• Offers a computationally efficient and superior alternative to existing methods for fetal brain analysis.