It is essential to accurately segment the fetal head and pubic symphysis in prenatal ultrasound images to assess labor progression and prevent complications. Although intrapartum transperineal ultrasound provides valuable insights into fetal head descent and pelvic position for clinicians, its practical application is limited by challenges such as noise, artifacts, and blurred boundaries. To address these challenges, this paper proposes a novel automatic segmentation method named Multi-Frequency Attention-UNeXt(MFA-UNeXt), based on a multi-frequency attention mechanism. The model utilizes discrete cosine transform to convert spatial domain information into frequency domain components, which are processed at both the channel and height levels to enhance feature extraction. At the channel level, the model increases sensitivity to different feature frequencies, improving fine feature extraction. At the height level, frequency-based processing captures both global structural information and fine details. This dual-level frequency attention mechanism synergistically enhances the model’s ability to detect features in complex scenarios. Additionally, MFA-UNeXt incorporates a pyramid pooling module to gather multi-scale contextual information and a shifted block module to expand the receptive field, further improving sensitivity to boundaries and intricate details. The model demonstrates outstanding performance, achieving an Intersection over Union of 94.72% and a Dice coefficient of 97.24% with only 0.26M parameters. It effectively segments complex structures, such as the fetal head and pubic symphysis, underscoring its substantial potential for clinical applications.