Progressive Semi-supervised Landmark Detection Algorithm For Intrapartum Ultrasound Measurement

Zelan Li, Hansen Zhang, Zhengyang Zhang, chengyan, Siqi Wang, Jianning Chi

Keywords: Landmark detection, Semi-supervised learning, Intrapartum ultrasound.

TL;DR: We propose a progressive semi-supervised landmark detection algorithm for intrapartum ultrasound AoP measurement. Following a coarse-to-fine strategy, We proposed a progressive learning strategy for learning landmark from area prior to landmark.

Abstract: The angle of progression (AoP) in intrapartum ultrasound is critical for evaluating fetal head descent and rotation during labor, and the angle formed by these three points (PS1, PS2, and FH1). Manual AoP measurement is time-consuming, labor-intensive, and lacks standardization—limitations. However, automated methods encounter two significant challenges: firstly, the scarcity of landmark annotations provided by experienced obstetricians may lead to network overfitting and poor generalization; secondly, the anatomical landmarks in ultrasound images are often too small, resulting in insufficient information and feature learning for algorithms. To address these challenges, inspired by the clinical workflow of manual AoP assessment, we propose a progressive semi-supervised landmark detection algorithm, which first locates and identifies the pubic symphysis (PS) and the fetal head (FH) region, and then detects the landmarks of three keypoints to calculate the AoP. Specifically, in the first stage, we utilize the spatial information of landmarks to generate scribbles of the foreground and background of the PS and the FH region. These scribbles are fed to a frozen segmentation foundation model named ScribblePrompt to get coarse segmentation and detection results as pseudo labels, which can help the network concentrate on PS and FH regions. After the first stage of pseudo-label pre-training, the following fine-tuning utilizes pre-trained models to learn landmarks with confidence-guided weight loss to train on labeled and unlabeled data, improving the robustness and generalization of the algorithm. The experimental results show that our algorithm achieved good landmark detection results.

Submission Number: 6