Towards contrast- and pathology-agnostic clinical fetal brain MRI segmentation using SynthSeg

Area of Science:

• Medical Imaging
• Neuroscience
• Artificial Intelligence

Background:

• Magnetic Resonance Imaging (MRI) is vital for fetal neurodevelopment research.
• Automated structural annotation of fetal brain MRIs using Deep Learning (DL) is crucial for quantitative analysis.
• Convolutional Neural Networks (CNNs) often struggle with domain shift, limiting performance on diverse subject data.

Purpose of the Study:

• To train DL networks for robust automatic segmentation of fetal brain MRIs across various domain shifts, including physiological and acquisition differences.
• To enhance domain generalizability of segmentation networks, especially for shape-based abnormalities in pathological cases.

Main Methods:

• Introduced a novel data-driven train-time sampling strategy to maximize training dataset diversity.
• Adapted the sampling strategy and integrated existing data augmentation techniques within the SynthSeg framework.
• Utilized domain randomization within SynthSeg to generate diverse training data for enhanced network generalization.

Main Results:

• Achieved notable improvements in segmentation quality for fetal brain MRIs with significant anatomical abnormalities (p < 1e-4).
• Observed a slight decrease in performance for cases with fewer abnormalities, indicating a trade-off in generalizability.
• Demonstrated the effectiveness of the data-driven sampling strategy in improving domain generalization.

Conclusions:

• The developed data-driven sampling strategy significantly enhances the domain generalizability of DL networks for fetal brain MRI segmentation.
• The approach shows particular promise for segmenting brains with pathological shape-based differences.
• This work provides a foundation for developing adaptive data-driven sampling strategies in other AI-driven medical imaging pipelines.