Standardizing Obstetric Ultrasound Segmentation Using Unpaired Domain Translation Techniques

Prenatal ultrasound is essential for fetal monitoring, yet access in low-resource settings remains limited by the shortage of trained personnel and variability in imaging equipment. Volume Sweep Imaging (VSI) enables acquisition by non-experts and provides diagnostically useful cine-loops for physician review. However, when extending VSI interpretation to artificial intelligence (AI) pipelines, cross-scanner variability introduces domain shift - scanner-dependent differences in contrast, speckle, and resolution that degrade model generalizability. In this study, we evaluate unpaired domain translation methods to standardize obstetric VSI across two scanners (Butterfly iQ+ and Mindray DP10). We implemented CycleGAN, denoising diffusion GANs, and a sequential (CG→Diff) approach. Performance was assessed with distribution similarity (PSNR, MI, BC, MAE) and structural similarity (SSIM, LNCC, CSS) metrics, alongside qualitative segmentation analysis. Results show that the hybrid method achieved the most balanced performance, improving PSNR (21.69) and LNCC (0.59) while preserving anatomical structures. These findings highlight the potential of adversarial-diffusion pipelines to mitigate domain shift and enable scalable AI-assisted obstetric ultrasound in low-resource environments.