Unlabeled Data-Driven Fetal Landmark Detection in Intrapartum Ultrasound

Chen Ma, Li Yunshu, Bowen Guo, Jing Jiao, Huang Yi FDU, Yuanyuan Wang, Yi Guo

Keywords: Intrapartum Ultrasound ; Angle of Progression; Keypoint Detection; Pseudo Labels; Model Pretraining

Abstract: The angle of progression (AoP) is a critical parameter for assessing fetal head descent during labor, requiring identification of three anatomical landmarks in intrapartum ultrasound images. Manual annotation is time-consuming and prone to inter- and intra-observer variability, while automated methods are hindered by limited labeled data and domain shifts across ultrasound devices. We propose an automated fetal biometry method for AoP calculation based on a modified TransUNet architecture with TinyViT backbone. The design integrates (i) MAE-assisted knowledge distillation from an Ultrasound Foundation Model (USFM) for robust representation learning, (ii) label perturbation to enhance robustness and cross-device generalization, and (iii) semi-supervised learning with pseudo-labeling to leverage unlabeled data. The network predicts heatmaps for landmark localization and calculates AoP from the detected coordinates. On the IUGC2025 Challenge test set, the proposed method achieved a mean radial error of 11.6749 pixels and a mean absolute AoP error of 3.8061 degrees.

Submission Number: 5