Comprehensive analysis of synthetic learning applied to neonatal brain MRI segmentation

Area of Science:

• Medical Imaging
• Neuroscience
• Artificial Intelligence

Background:

• Neonatal brain MRI segmentation is challenging due to anatomical variability and signal intensity changes during gestation.
• Existing segmentation techniques struggle with variations in image contrast and anatomical configurations.

Purpose of the Study:

• To evaluate synthetic learning, a contrast-independent model, for segmenting neonatal brain MRI.
• To assess the robustness of synthetic learning against image contrast variations and identify its limitations.

Main Methods:

• Experiments utilized the developmental Human Connectome Project dataset (over 700 neonates, 26-45 weeks postconception).
• A standard UNet was trained on limited data, and its performance was compared to a synthetic learning model.
• The synthetic training set was enhanced with motion artifacts and white matter over-segmentation.

Main Results:

• Standard UNet models learned domain-specific intensity features, while synthetic learning demonstrated robustness to contrast variations.
• Model performance was influenced by neonatal age; enrichment of the synthetic set improved results.
• Synthetic learning models can avoid systematic biases present in real training data and ground truth.

Conclusions:

• Synthetic learning is an effective and robust method for neonatal brain MRI segmentation.
• An adapted synthetic learning approach offers advantages for multisite studies and clinical applications by mitigating data variability and biases.