Cardiovascular diseases are a leading cause of mortality worldwide, necessitating accurate segmentation of cardiac structures from 3D CT and MR images for effective diagnosis and treatment planning. Manual segmentation is time-consuming, labor-intensive, and challenging due to the complexity of cardiac anatomy structures. General CNNs, while effective, often struggle to generalize across different imaging domains. To address these issues, we developed a PPM-UNet, which integrates a probabilistic perturbation module (PPM) into the traditional U-Net architecture. Here, the PPM is introduced to control randomness during training, enhancing the model’s robustness and improving its performance across different imaging domains. Furthermore, the PPM is lightweight and easy to implement, making it a practical addition to existing network architectures for domain generalization tasks in medical imaging. Evaluated on the WHS++ dataset, our method demonstrates superior performance in segmenting cardiac structures, particularly in unseen domains, as evidenced by improved Dice Similarity Coefficient, Precision, Sensitivity, and Hausdorff Distance metrics. In addition to its performance benefits, the proposed module requires minimal computational overhead, making it practical for real-world clinical applications.