A longitudinally-consistent deep framework for joint subcortical segmentation and registration of infant brains

Area of Science:

• Neuroimaging
• Developmental Neuroscience
• Medical Image Analysis

Background:

• Infant brain MRI segmentation is vital for neurodevelopment studies but challenging due to rapid growth and poor contrast.
• Existing methods often segment scans independently, leading to inconsistencies over time.

Purpose of the Study:

• To develop a novel deep learning framework, C2FSRnet, for longitudinally consistent joint subcortical segmentation and registration of infant brain MR images.
• To improve the accuracy and reliability of neurodevelopmental studies by addressing limitations of current methods.

Main Methods:

• Proposed C2FSRnet, a coarse-to-fine network integrating segmentation and registration for longitudinal infant brain MRI.
• Employed a joint encoder in the coarse stage and incorporated signed distance maps in the fine stage for enhanced feature learning and spatial guidance.
• Trained the framework using multiple longitudinal scans per subject to learn within-subject anatomical correspondences.

Main Results:

• C2FSRnet demonstrated superior accuracy and longitudinal consistency compared to 15 state-of-the-art methods on the Baby Connectome Project and developing Human Connectome Project datasets.
• The framework successfully performed joint affine/deformable registration and subcortical segmentation.
• Achieved robust performance in characterizing early subcortical development in both healthy and clinical populations.

Conclusions:

• C2FSRnet offers a robust solution for joint segmentation and registration of longitudinal infant brain MRIs.
• The proposed method enhances the characterization of early neurodevelopmental trajectories.
• This framework holds significant potential for both research and clinical applications in pediatric neurology.