Evaluating Synthetic Data Generation for Domain Generalization in Fetal Brain MRI Segmentation
Researchers have developed FetalSynthSeg, a novel framework for generating synthetic fetal brain MRI data to improve segmentation accuracy and domain generalization. The study found that simple Gaussian mixture-based intensity modeling and intensity clustering were more effective than complex physics-based simulations for enhancing out-of-domain robustness. FetalSynthSeg achieved state-of-the-art performance on FeTA 2024 datasets and demonstrated robust segmentation capabilities across different modalities and sites, outperforming existing methods like BOUNTI and nn-Net. AI