Accurate segmentation of cervical and fetal head structures in ultrasound imaging is critical for assessing preterm birth risk and monitoring labor progression. However, ultrasound images are often affected by speckle noise, low contrast, and anatomical variability. Existing fully supervised AI solutions require large, annotated datasets and may struggle with robustness, limiting scalability and real-world clinical adoption.
OASIS-Net introduces a dual-space adversarial semi-supervised framework that delivers high-precision segmentation using a small fraction of labeled data. By combining input-level and model-level perturbation strategies within a unified training approach, it improves robustness and boundary accuracy while maintaining near real-time performance. The system enables automated cervical length and angle-of-progression measurements, supporting objective obstetric screening and intrapartum decision-making.
Framework diagram illustrating the end-to-end training workflow of OASIS-Net, including supervised and unsupervised branches with dual-space adversarial perturbations.