Background: The accurate estimation of labor progress parameters by means of intrapartum transperineal ultrasound is of utmost significance for pregnant patients. It allows for the timely identification of deviations from the norm, thereby facilitating prompt intervention to minimize or avert potential maternal or fetal problems. The majority of existing methods hinge upon structural information and morphological prior knowledge. However, the parameter estimation achieved through these methods is characterized by limited accuracy and continues to pose significant challenges. Results: In this research, we put forward a novel multi-path refinement U-Net that combines recurrent residuals and attention (MRU-CRRA). To enhance the performance of segmentation and estimation, we comprehensively take into account the integrity and continuity of the boundaries of the fetal head and the pubic symphysis. MRU-CRRA makes use of local neighboring features as well as global multi-scale high-level semantic information. Under the guidance of cross-attention, both the spatial proximity and the feature similarity are effectively integrated.