Unified Model for Children's Brain Image Segmentation With Co-Registration Framework Guided by Longitudinal MRI

Area of Science:

• Medical imaging analysis
• Pediatric neuroimaging
• Computational neuroscience

Background:

• Accurate segmentation of pediatric brain structures is vital for longitudinal studies.
• Existing methods struggle with longitudinal data due to annotation scarcity and tissue variability, leading to inconsistent results.
• Current approaches often require multiple MRI sequences, limiting their applicability.

Purpose of the Study:

• To develop a unified model for segmenting brain images in children from neonates to preschoolers.
• To address the challenges of limited annotated data and dynamic tissue intensity variations in longitudinal pediatric neuroimaging.
• To enable accurate analysis of brain development trends across different time points.

Main Methods:

• A two-stage approach incorporating a co-registration framework for gold-standard segmentation guided by longitudinal data.
• Development of a unified segmentation model, SE-VB-Net, a convolutional network combining VB-Net with Squeeze-and-Excitation blocks.
• The model accommodates both T1- and T2-weighted MRI, and notably, a single T1-weighted image as input.

Main Results:

• The proposed SE-VB-Net method achieved high performance (>92%) on longitudinal pediatric brain MRI segmentation.
• Demonstrated suitability for analyzing images with significant appearance variations, outperforming single time-point methods.
• Validation on a large dataset (320 subjects) and two external datasets confirmed robustness.

Conclusions:

• The novel unified model and co-registration framework significantly improve the accuracy of longitudinal brain segmentation in children.
• SE-VB-Net offers a versatile and robust solution for pediatric neuroimaging analysis, broadening its application scope.
• This method facilitates more reliable estimation of brain development trajectories in early childhood.