Accurate assessment of labor progress is vital for reducing unnecessary cesarean sections and associated risks. Intrapartum ultrasound imaging provides essential insights into fetal positioning, particularly of the fetal head (FH) and pubic symphysis (PS), which are crucial for calculating the angle of progression (AoP) and head-symphysis distance (HSD). These biomarkers allow the clinician to decide on the mode of fetal delivery. However, analyzing these images is challenging due to high variability and limited labeled data, especially because imaging needs to be performed in a stressful labor and delivery environment. In this study, we propose a multi-task deep learning solution to 1) classify clinically acceptable standard planes in ultrasound scans and 2) locate and segment the FH and PS, which allow for the computation of AoP and HSD. This work is part of the MICCAI IUGC 2024 challenge. The method is based on a multi-task U-Net model that has been trained in a weakly-supervised approach through the generation of synthetic masks. To optimize the learning of multiple tasks, we introduce a dynamic loss scaling approach. In addition, we investigate complex data augmentation strategies to mitigate the overfitting problem. Our experiments demonstrate that incorporating synthetic labels and dynamic loss scaling significantly improves model performance. Specifically, our method outperforms the baseline by achieving a 14% increase in classification accuracy and a 6.3% improvement in segmentation Dice score on the test set. We also observe reduced deviations in AoP and HSD measurements, enhancing clinical assessment reliability. These results indicate that our multi-task learning framework effectively enhances the analysis of intrapartum ultrasound images, potentially aiding clinicians in better assessing labor progress and making informed decisions. The code is publicly available at t.ly/ze6gN .