Evaluating Synthetic Data Generation for Domain Generalization in Fetal Brain MRI Segmentation

Technical Summary: Evaluating Synthetic Data Generation for Domain Generalization in Fetal Brain MRI Segmentation

Problem Statement
Fetal brain tissue segmentation from Magnetic Resonance Imaging (MRI) is essential for studying neurodevelopment and detecting abnormalities. However, automated analysis is hindered by severe domain shifts arising from variations in scanner hardware, field strengths (ranging from 0.55T to 3T), acquisition protocols, super-resolution (SR) reconstruction methods, and gestational age. These factors create distributional mismatches between training and testing data, leading to poor generalization. Furthermore, the scarcity of annotated fetal data and the high cost of manual segmentation (often taking hours per case) limit the availability of large, diverse training sets. While Single-Source Domain Generalization (SSDG) aims to train models on a single source domain to generalize to unseen targets, existing methods often struggle with the unique morphological changes and limited tissue contrast inherent to fetal imaging.

Methodology
The authors investigate the efficacy of Domain Randomization (DR) for SSDG, specifically comparing physics-based simulations against intensity-based randomization strategies. The study focuses on the FetalSynthSeg framework, an extension of SynthSeg tailored for fetal MRI.

1. Synthetic Data Generation Strategies:

   • Physics-Based (FaBiAN): Uses numerical phantoms and extended phase graph (EPG) formalism to simulate T2-weighted images based on literature T1/T2 values.
   • Randomized Physics (randFaBiAN): Similar to FaBiAN but samples T1/T2 values from broad, unconstrained distributions.
   • Gaussian Mixture Model (GMM) Randomization (SynthSeg): Generates synthetic images from label maps using GMMs with randomized intensity parameters, bias fields, and artifacts, without relying on physical signal modeling.
   • FetalSynthSeg: An enhanced GMM approach that addresses the coarse nature of fetal annotations (7 classes) by:
     • Grouping labels into four meta-classes (White Matter, Gray Matter, CSF, non-brain).
     • Applying unsupervised intensity-based clustering (Expectation-Maximization) within these meta-classes to create sub-classes.
     • Sampling independent GMM parameters for each sub-class to model intra-class heterogeneity.

2. Training and Baselines:

   • All models utilize a standard 3D U-Net backbone trained with a combined Dice + Cross-Entropy loss.
   • Baselines: The study compares FetalSynthSeg against:
     • FetalRealSeg: Trained exclusively on real images with standard augmentation.
     • RealSynthHybrid: A 50/50 mix of real and synthetic data.
     • State-of-the-Art (SoTA): BOUNTI, the FeTA 2024 winning method (cesne-digair), and an nnU-Net ensemble.

3. Experimental Design:

   • Experiments span 348 fetal subjects from four sites (KISPI, VIEN, CHUV, KCL) and the dHCP dataset, covering T1w and T2w contrasts and various field strengths.
   • Evaluations include cross-domain testing (training on one site/SR method, testing on others), ablation studies on intensity clustering, and prospective validation on multi-echo time (TE) data for T2 mapping.

Key Contributions

1. Comparison of Generation Strategies: The paper systematically evaluates physics-based vs. randomization-based intensity generation, demonstrating that simple GMM-based modeling outperforms complex physics-based simulations for domain generalization.
2. FetalSynthSeg Framework: The authors propose a method that combines meta-label grouping with intensity sub-clustering. This approach captures the heterogeneity of fetal tissues despite coarse annotations, significantly improving Out-of-Domain (OOD) robustness.
3. Robustness Across Modalities: The study demonstrates that FetalSynthSeg achieves state-of-the-art performance on T2w datasets while, for the first time, enabling robust segmentation of T1-weighted fetal brain images (80 Dice score on dHCP-T1w) using a single network trained exclusively on synthetic T2w data.
4. Prospective Validation: The method is validated on prospectively acquired multi-TE data at 0.55T, showing consistent segmentation across echo times where other methods fail.

Results

• Performance: FetalSynthSeg achieved a global mean Dice score of 74.9 in cross-domain experiments, outperforming FaBiAN (63.3), randFaBiAN (68.3), and standard SynthSeg (72.2).
• Ablation: Intensity sub-clustering was found to be critical; performance gains saturated at approximately six subclasses. The meta-class strategy provided an additional ~2 Dice point improvement over standard SynthSeg.
• OOD Generalization: On the dHCP T1w dataset (a significant domain shift), FetalSynthSeg achieved a Dice score of 80.1%, whereas all other methods (including those trained on real T2w data) dropped below 25%.
• In-Domain Performance: On T2w datasets close to the training distribution, FetalSynthSeg remained competitive with or slightly below models trained on real data (e.g., FetalRealSeg), confirming that synthetic training does not necessarily sacrifice in-domain accuracy.
• Failure Modes: The model showed reduced performance in severely pathological cases (e.g., ventriculomegaly) and older gestational ages, often due to annotation limitations or anatomical changes outside the training distribution.

Significance
The paper claims that structured intensity randomization is a more effective strategy for SSDG in fetal MRI than physically grounded simulations. By decoupling the training data generation from specific acquisition protocols and focusing on broad, randomized intensity distributions, FetalSynthSeg creates a contrast-agnostic model. This enables reliable segmentation across diverse field strengths, reconstruction methods, and modalities (T1w and T2w) without requiring target domain annotations. The authors posit that this approach provides a practical, efficient path toward robust fetal brain segmentation in data-scarce environments, potentially supporting downstream tasks such as T2 mapping and inter-modality registration. The code and models are made publicly available to facilitate further research.