Fetal Cortical Plate Segmentation Using Fully Convolutional Networks With Multiple Plane Aggregation.

Area of Science:

• Neuroimaging
• Developmental Neuroscience
• Medical Image Analysis

Background:

• Fetal magnetic resonance imaging (MRI) offers in vivo insights into human brain development.
• Accurate segmentation of the fetal cortical plate (CP) is essential for quantitative analysis of brain development, including volume and folding.
• Current segmentation methods may lack the accuracy and automation required for reliable analysis.

Purpose of the Study:

• To develop a fully convolutional neural network for automatic segmentation of the fetal cortical plate (CP).
• To improve segmentation accuracy using a novel hybrid loss function and 3D information.
• To enable reliable and automatic quantification of fetal brain structures.

Main Methods:

• Proposed a fully convolutional neural network for automatic CP segmentation.
• Developed a novel hybrid loss function to enhance segmentation precision.
• Employed multi-view aggregation (axial, coronal, sagittal) and test-time augmentation for robust 3D analysis.

Main Results:

• Achieved high Dice coefficients (0.907 ± 0.027 left, 0.906 ± 0.031 right) and low mean surface distance errors (0.182 ± 0.058 mm left, 0.185 ± 0.069 mm right).
• Demonstrated strong correlation (R ² > 0.941) between automated and manual segmentations for CP volume, surface area, and global mean curvature.
• Confirmed significant improvements in segmentation accuracy attributed to the hybrid loss function and multi-view/test-time augmentation strategies.

Conclusions:

• The proposed automated segmentation method accurately quantifies fetal cortical structures.
• The novel hybrid loss function and 3D analysis techniques enhance segmentation reliability.
• This method provides a valuable tool for advancing the study of fetal brain development.