In radiation oncology, there is emerging evidence showing that cardiotoxicity may related to the radiation dose to cardiac substructures. Manual delineation of cardiac substructures in non-contrast CT is tedious and challenging. In this work, we utilized a deep learning-based approach to achieve auto-delineation on nine cardiac substructures. Model performances with different network designs and training strategies were studied and compared. To evaluate the clinical acceptability, our model was tested on both in-distribution data and two types of clinically common outlier data.