Accurate obstetric ultrasound segmentation is hampered by speckle noise and scarce annotations. We propose OASIS-Net, a dual-space adversarial semi-supervised framework that trains a single DeepLabV3$+$ backbone by minimizing one unified consistency loss. The loss couples input-space adversaries (iterative FGSM with $K=3$ steps, $\epsilon =4/255$) and weight-space gradient-aligned perturbations (DGAP, weight scale $=0.5$) whose influence grows with a sigmoid ramp ($T_{\text{ramp}}=20$, $\alpha _{\max }=1.0$). Pseudo-labels are accepted with a confidence threshold of 0.95 and the unlabeled loss weight is 1.0. We evaluate OASIS-Net on two public obstetric benchmarks: FUGC (50 labeled, 450 unlabeled) and PSFH (5,101 frames, 70% unlabeled). Using 20% of labels, the method attains Dice = 96.53% and HD$_{95}$ = 3.86 px on FUGC, and Dice = 97.16% and HD$_{95}$ = 2.34 px on PSFH. Ablation shows that removing either perturbation stream reduces Dice by up to 1.8 percentage points. The trained model runs at 18.96 frames s$^{-1}$ on a single RTX 4060 Ti and produces high-precision masks that enable automated cervical-length and angle-of-progression measurements for objective obstetric screening and intrapartum monitoring. These results demonstrate that jointly enforcing input- and parameter-space adversarial consistency yields a label-efficient, robust solution for obstetric ultrasound segmentation and supports real-time clinical use.