Structure-Aware Perturbation Consistency Framework for Semi-Supervised Fetal Ultrasound Image Segmentation

Accurate measurement of the Angle of Progression (AoP), derived from precise segmentation of the pubic symphysis (PS) and fetal head (FH) in intrapartum ultrasound images, is essential for labor management. While semi-supervised learning addresses annotation scarcity through consistency regularization, existing methods suffer from three limitations: they enforce globally uniform constraints, ignoring that prediction uncertainty concentrates at anatomical boundaries; standard loss functions fail to handle severe size disparity between anatomical structures; and sole reliance on teacher-generated pseudo-labels causes the model to learn from unreliable targets, leading to error accumulation. We argue that effective semi-supervised segmentation requires structure-aware adaptation to local reliability. Based on this insight, we propose Structure-Aware Perturbation Consistency (SAPC) with three complementary mechanisms. Specifically, Structure-Aware Uncertainty Correction (SAUC) identifies uncertain regions through inter-student disagreement and enforces selective consistency via mutual calibration. Furthermore, Adaptive Small Region Loss (ASRL) employs hierarchical class-boundary-instance reweighting to amplify supervision for smaller structures. Additionally, Uncertainty-aware Pseudo-label Filtering (UPF) integrates teacher entropy with structural disagreement to suppress unreliable pseudo-labels. Experiments on the PSFHS dataset demonstrate AoP errors of $8.53^\circ$ and $6.10^\circ$ with 10% and 20% labeled data, respectively, while cross-modality validation on LA and BraTS19 further shows the generalization potential of SAPC.