A Two-Stage Semi-Supervised Ensemble Framework for Automated Angle of Progression Measurement in Intrapartum Ultrasound

Bo Deng, Yu Chen, Zilun Peng

Keywords: Intrapartum Ultrasound, Angle of Progression (AoP), Two-Stage Model, Semi-Supervised Learning

TL;DR: We introduce a two-stage semi-supervised ensemble framework that combines coarse localization and fine-grained refinement to achieve high accuracy in automated Angle of Progression (AoP) measurement.

Abstract: Accurate and reproducible measurement of the Angle of Progression (AoP) from intrapartum ultrasound is critical for modern labor management, yet manual annotation is hindered by significant intra- and inter-observer variability and workflow inefficiencies. To address this, we propose a fully automated, two-stage deep learning pipeline for precise landmark localization. The first stage employs a multi-model ensemble of U-Net architectures with diverse backbones (EfficientNet-B4 and -B7), trained under a Mean Teacher semi-supervised framework to leverage both labeled and unlabeled data. This stage generates a robust coarse prediction by performing a per-keypoint weighted average of the fused heatmaps. In the second stage, a dedicated Res-Net-18-based regression model refines the position of each landmark by predicting a precise offset from its coarse location on a localized image patch. Our integrated approach, trained on a combined dataset from the 2024 and 2025 IUGC challenges, demonstrates highly competitive performance, achieving a Mean Radial Error (MRE) of 12.7888 pixels and a mean Absolute Parameter Difference (APD) of 4.4581 degrees for the AoP on the test set. This automated framework promises to enhance diagnostic consistency and streamline clinical workflows, aligning with the WHO's vision for improved intrapartum care.

Submission Number: 4