Intrapartum ultrasound monitoring is critical for maternal-fetal safety, yet traditional manual annotation of key anatomical landmarks (PS1, PS2, FH1) faces bottlenecks such as significant inter-observer variability and time-intensive processes, hindering standardized implementation of the WHO Labor Care Guide (LCG). This study proposes DSNT-DeepUNet, a deep learning-based ultrasound coordinate prediction model. By integrating a U-Net backbone with a Differentiable Spatial to Numerical Transform (DSNT) layer, it achieves end-to-end mapping from raw ultrasound images to keypoint coordinates. The model employs a multi-task loss function to simultaneously optimize coordinate accuracy and heatmap distribution, while an 8-fold cross-validation strategy and dynamic data augmentation techniques significantly enhance generalization capability. On an independent test set, the model achieved an angle of progression prediction error of 4.7005 pixels and an average point distance error of 14.7712 pixels, with PS1 and PS2 localization errors at 9.0600 and 11.5661 pixels respectively, ranking sixth in a public challenge. This solution successfully eliminates subjective variations in manual annotation, demonstrating effective and precise ultrasound coordinate prediction.