Semi-supervised Fetal Brain Parcellation via Hierarchical Learning Framework.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Developmental Neuroscience

Background:

• Automatic fetal brain region segmentation using MRI is crucial for studying prenatal development.
• Manual segmentation is labor-intensive and limited by scarce annotated data.
• Existing methods often overlook the hierarchical structure and inter-regional relationships in the brain.

Purpose of the Study:

• To introduce a novel hierarchical segmentation method for detailed fetal brain parcellation into 87 regions.
• To address the challenges of limited annotated data and improve segmentation robustness.
• To enhance the understanding of prenatal brain growth and development.

Main Methods:

• A three-level coarse-to-fine network architecture is proposed for hierarchical segmentation.
• A data augmentation module simulates imaging variations to enhance robustness.
• Semi-supervised learning combines simulated and limited real labeled data for training.

Main Results:

• The method achieved a high Dice score of 91.42% on fetal brain MR images.
• It outperformed the leading nnUNet method (88.77%) in segmentation accuracy.
• Demonstrated robust performance across diverse imaging conditions and scanner variability.

Conclusions:

• The proposed hierarchical segmentation method effectively addresses limitations in current fetal brain parcellation.
• It offers a robust and accurate approach for studying prenatal brain development using MRI.
• This technique has the potential to advance research in developmental neuroscience and clinical applications.