Synthetic medical images can augment training datasets where acquisition is costly or ethically constrained. In MRI imaging, subtle anatomical inconsistencies in synthetic images may reduce model generalisation if they deviate from plausible anatomy. We introduce SynthUterus ROI, a dataset of 800 region-of-interest (ROI)-cropped 96x96 synthetic uterus MRI images generated with diffusion models conditioned on uterine orientation. All images were independently annotated by clinicians for visual realism, orientation correctness and anomalies. SynthUterus ROI is designed for training and validation of (normative) medical imaging models, and is the first dataset to label uterus orientation explicitly. Publicly available at Zenodo (10.5281/zenodo.18297879).
Müller et al. (Mon,) studied this question.