Segment Anything Model for fetal head-pubic symphysis segmentation in intrapartum ultrasound image analysis

Abstract

The Angle of Progression (AoP), determined by the contour delineations of the pubic symphysis and fetal head (PSFH) in intrapartum ultrasound (US) images, is crucial for predicting delivery mode and significantly influences labor outcomes. However, automating AoP measurement based on PSFH segmentation remains challenging due to poor foreground-background contrast, blurred boundaries, and anatomical variability in shapes, sizes, and positions during labor. We propose a novel Segment Anything Model (SAM) framework, AoP-SAM, designed to enhance the PSFH segmentation, AoP measurement and outcome prediction, tackling the challenges of small target segmentation and accurate boundary segmentation. It synergistically combines CNNs and Transformers within a cooperative encoder to process complex spatial relationships and contextual information to segment the PSFH. In this encoder, we devise a multi-scale CNN branch to capture intrinsic local details, which compensates for the defects of the Transformer branch in extracting local features. Further, a cross-branch attention module is applied to improve prediction by fostering the effective information exchange and integration between two branches. Evaluations on the benchmark dataset demonstrate that our method achieves state-of-the-art (SOTA) performance. Specifically, in the PSFH segmentation task, the AoP measurement task, and the AoP classification task, we achieved a DSC of 0.8745 on the PS structure, a $\Delta$AoP of 7.6743°, and an F1-score of 0.7719, respectively. Compared to the second-ranking method, these results represent improvements of 2.5%, 6.3%, and 1.1%. Our study presents a framework for intrapartum biometry, offering significant advancements in labor monitoring and delivery mode prediction in clinical settings. Future efforts will focus on optimizing AoP-SAM for resource-constrained environments, highlighting its potential for lightweight adaptation. https://github.com/maskoffs/AoP-SAM.