Improving cross-domain brain tissue segmentation in fetal MRI with synthetic data

TL;DR

FetalSynthSeg is introduced, a domain randomization method to segment fetal brain MRI, inspired by SynthSeg, and shows that models trained solely on synthetic data outperform models trained on real data in out-ofdomain settings, validated on a 120-subject cross-domain dataset.

Abstract

Segmentation of fetal brain tissue from magnetic resonance imaging (MRI) plays a crucial role in the study of in utero neurodevelopment. However, automated tools face substantial domain shift challenges as they must be robust to highly heterogeneous clinical data, often limited in numbers and lacking annotations. Indeed, high variability of the fetal brain morphology, MRI acquisition parameters, and superresolution reconstruction (SR) algorithms adversely affect the model's performance when evaluated out-of-domain. In this work, we introduce FetalSynthSeg, a domain randomization method to segment fetal brain MRI, inspired by SynthSeg. Our results show that models trained solely on synthetic data outperform models trained on real data in out-ofdomain settings, validated on a 120-subject cross-domain dataset. Furthermore, we extend our evaluation to 40 subjects acquired using lowfield (0.55T) MRI and reconstructed with novel SR models, showcasing robustness across different magnetic field strengths and SR algorithms. Leveraging a generative synthetic approach, we tackle the domain shift problem in fetal brain MRI and offer compelling prospects for applications in fields with limited and highly heterogeneous data.

Figures (8)

• Figure 1: Domain shifts across data splits in fetal SR MRI. (GA in weeks, site-SR). A & C - pathological, B & D - neurotypical.
• Figure 2: Synthetic image generation framework. Original segmentation labels are merged to create a 4-meta-label tissue map (CSF, WM, GM, skull). EM clustering then divides each meta-label into 1 to 4 subclasses, capturing tissue heterogeneity. A generative model uses these split meta-labels to produce synthetic images.
• Figure 3: Comparison of the out-of-distribution performance of the segmentation models (mdsc - mean Dice score across all tissues). (A)baseline (light) vs FetalSynthSeg (dark). Data split: KISPI-mial (green), KISPI-irtk (red), CHUV-mial (blue). See Figure \ref{['fig:segm_all']} from the appendix for a comparison with in-distribution performance as well as results split by gestational age. The dashed horizontal line inside the boxplot corresponds to the mean value. (B) Pathological vs neurotypical, aggregated across all models.
• Figure 4: Cross-domain model inference qualitative results. (A) Model trained on CHUV-mial and tested on KISPI-irtk/KISPI-mial. (B) Model trained on KISPI-irtk and tested on KCL-svrtk/KCL-nesvor.
• Figure 5: Segmented tissue volumes vs GA for KCL data reconstructed with SVRTK (top row) and NeSVoR (bottom row). KISPI_irtk_base (blue) and KISPI_irtk_synth (orange) model predictions are compared to FeTA reference values (green) which are based on the ground truth segmentation of 40 healthy subjects selected across all splits. Lines are second-order polynomial fit and a corresponding shaded area is a confidence interval. See Figure \ref{['fig:normALL']} in the Supplement for all tissues evaluation.