CNN-stylized dual-path CNN-Transformer for fetal head and pubic symphysis segmentation in intrapartum ultrasound

In intrapartum ultrasound (IU) imaging, reliable segmentation of the fetal head (FH) and the pubic symphysis (PS) is a prerequisite for automated quantification of the angle of progression (AoP), a clinically meaningful indicator for predicting delivery outcomes and mitigating maternal-neonatal complications. However, existing CNN-Transformer hybrid models often suffer from unstable attention learning under limited annotated medical data, which can lead to attention collapse. In addition, they tend to underexploit boundary cues in IU images, where contours are frequently degraded by noise and artifacts. To address these issues, we propose a CNN-stylized dualpath CNN-Transformer framework tailored for IU segmentation. The encoder consists of a CNN branch and a CNN-stylized Transformer branch to balance local detail representation and long-range dependency modeling while improving the robustness of attention learning. A Transformer-to-CNN fusion module is further introduced to enhance inter-branch interaction and feature complementarity. In the decoder, a Reverse-Additive Boundary Refinement module explicitly models the foreground-background transition region, progressively refining boundary representations. Extensive experiments on three datasets demonstrate that the proposed method consistently achieves superior overall segmentation accuracy and improved boundary quality compared with state-of-the-art approaches.