FetalSynthSeg beats domain shifts in fetal brain MRI segmentation

Fetal brain tissue segmentation from MRI is essential for studying neurodevelopment but suffers from data heterogeneity and limited annotations. Domain randomization has emerged as a promising strategy for single-source domain generalization by synthesizing training images with randomized artifacts, contrast, and resolution. In this work, a team of 12 researchers introduces FetalSynthSeg, a framework that evaluates several synthetic data generation strategies. They find that simple Gaussian mixture-based intensity modeling outperforms complex physics-based simulations, and that intensity clustering—subdividing tissue classes by intensity—improves out-of-domain (OOD) robustness.

Evaluated on 348 fetal subjects from four sites spanning 0.55–3T field strengths and both T1w and T2w contrasts, FetalSynthSeg achieves state-of-the-art performance on several FeTA 2024 testing datasets (80–85 Dice score). For the first time, it offers robust segmentation on non-T2w modalities, reaching 80 Dice on the dHCP-T1w dataset. Compared with leading methods like BOUNTI, nnU-Net ensemble, and the FeTA 2024 winner, FetalSynthSeg delivers comparable or superior accuracy while maintaining strong robustness across domain shifts. The code, model weights, and a Docker image for easy inference are publicly available.