Harmonized Segmentation of Neonatal Brain MRI

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neonatal Neuroscience

Background:

• Deep learning models for medical image segmentation often fail with data from different sources due to distribution shifts.
• Unsupervised domain adaptation (UDA) offers a solution by aligning data distributions without target labels.
• Accurate segmentation is crucial for analyzing neonatal brain development and conditions.

Purpose of the Study:

• To adapt deep learning segmentation models for T2-weighted MRI in preterm neonates using UDA.
• To evaluate and compare two UDA techniques for this specific clinical challenge.
• To assess the impact of UDA on tissue segmentation, cortical thickness, and clinical outcome associations.

Main Methods:

• Investigated two unsupervised domain adaptation techniques for T2-weighted neonatal brain MRI segmentation.
• Compared UDA methods against a fully-supervised baseline segmentation network.
• Analyzed tissue volumes and cortical thickness on harmonized data, controlling for age.

Main Results:

• UDA methods demonstrated potential in harmonizing MRI data from dissimilar sources.
• Cortical gray matter maps were successfully generated for preterm neonates.
• Preliminary analysis showed associations between cortical thickness and language outcomes.

Conclusions:

• Unsupervised domain adaptation is a viable strategy to improve neonatal brain MRI segmentation across different data distributions.
• Harmonized segmentation maps enable more robust analysis of brain development and clinical associations in preterm infants.
• This approach holds promise for advancing precision medicine in neonatal care.