Prostate cancer is one of the most prevalent malignancies among men, making tumor segmentation a crucial task for supporting diagnosis and treatment planning. However, the segmentation process is often time-consuming and requires substantial medical expertise, particularly in the context of an increasing number of cases combined with limited human resources and medical equipment, which leads to pressure and overload in diagnosis and treatment workflows. Moreover, data imbalance between tumor regions and healthy tissue or other anatomical structures, along with the challenge of effectively learning and leveraging intra-slice and inter-slice spatial information, often results in suboptimal segmentation performance. In this study, we develop and compare U-Net architectures with four different backbones (CNN, VGG, ResNet, and EfficientNet) using the Public Training and Development Dataset from the PI-CAI challenge for prostate tumor segmentation. The models are trained using a 3-fold cross-validation setup and evaluated with multiple metrics, including mAP, Dice, IoU, Precision, Recall, and F1-score. Our experimental results show that EfficientNet-based U-Net achieves the highest performance, with the mAP of 0.6667 averaged across three folds, outperforming the CNN and VGG variants by more than 23%. These findings demonstrate the potential of modern backbone architectures in enhancing prostate tumor segmentation accuracy, providing a promising direction for improving computer-aided diagnosis systems in clinical practice.