Researchers including Bjoern Menze from the Menze Lab at the Department of Quantitative Biomedicine have developed a novel training strategy to improve the robustness of automated fetal brain MRI segmentation across contrast variations and severe pathologies. The work extends the SynthSeg framework with a data-driven sampling approach that explicitly balances anatomical shape variability during training.
By prioritizing underrepresented and pathological brain morphologies and combining this with structure-specific synthetic deformations, the method significantly improves segmentation performance in cases with pronounced ventriculomegaly. At the same time, it maintains stable performance across a wide range of imaging domains, scanners, and reconstruction pipelines.
The study demonstrates how principled sampling strategies can enhance the clinical reliability of deep learning models for fetal neuroimaging and supports broader adoption of quantitative fetal MRI in research and clinical practice.
DOI: https://doi.org/10.1016/j.neuroimage.2026.121729