Accurate segmentation of fetal abdominal structures in ultrasound images is essential for prenatal diagnostic and fetal biometric evaluations. However, fetal ultrasound is often degraded by low contrast, speckle noise, and structural variability, making automatic segmentation challenging. This study presents a novel model called X-DSAU-Net, an enhanced Attention U-Net architecture with deep supervision, for the segmentation of key fetal abdominal structures - liver, stomach, vein, and artery - from 2D ultrasound images. Attention gates embedded in the skip connections focus on informative features while suppressing irrelevant activations. Deep supervision is applied at multiple decoder stages to guide intermediate layers, promote multi-scale feature learning, and improve convergence. Additionally, explainable artificial intelligence techniques, such as Grad-CAM, are integrated to visualize regions influencing the model’s predictions, enhancing interpretability. The model is trained and evaluated on a dataset of 1,588 annotated ultrasound images. On the test set, it achieves a mean Dice coefficient of 0.8581, mean Jaccard index of 0.7601, and Hausdorff distance of 7.01 pixels. Class-wise analysis reveals the highest Dice scores of 0.8904 and 0.8322 for liver and stomach respectively, with satisfactory performance on the vein and artery with Dice scores of 0.8023 and 0.7721 respectively. These results show that the model effectively segments both prominent and subtle structures, making it useful for fetal biometric assessment.