GMANet: Gate mamba attention for fetal head and pubic symphysis segmentation in ultrasound images analysis

Abstract

Accurate segmentation of fetal heads and the pubic symphysis (PSFH) in ultrasound images during childbirth is crucial for precise angle of progression (AoP) measurements, which enables clinicians to manage dystocia complications effectively. Conventional approaches relying on sonographer-dependent manual selection prove time-consuming and operator-sensitive, while concurrently coping with inherent ultrasound noise, anatomical occlusions, and substantial target shape or location variations. To overcome these challenges in small-target segmentation and boundary delineation, we present GMANet, a novel Mamba-based architecture. Our core design introduces the Gate Mamba Attention (GMA) that synergistically integrates selective state-space modeling with a gating mechanism, where sequence-aware attention of Mamba dynamically focuses on crucial spatial dependencies. At the same time, the fixed-parameter architecture maintains stable local feature extraction. Then we develop an Adaptive Pyramid Pooling Module (APPM) that enhances multiscale discriminability through parallel multi-depth pooling, effectively handling significant size disparities in medical targets. Subsequent feature refinement employs our Efficient Multiscale Attention (EMA) to aggregate multi-receptive-field context through parameter-efficient spatial-channel interactions adaptively. Finally, the proposed GMANet demonstrates statistically significant advantages when benchmarked against contemporary state-of-the-art (SOTA) segmentation methodologies on the PSFH dataset, achieving a composite score of 0.9326, F1-score of 76.04, and ΔAoP of 7.70°. This advancement holds significant promise for automating fetal imaging analysis, potentially improving clinical consistency while reducing operator dependence. Our code is available at https://github.com/AgamLi/GMANet-Gate-Mamba-Attention