Deep convolutional neural network (DCNN) has shown great success in various medical image segmentation tasks, including organ-at-risk (OAR) segmentation from computed tomography (CT) images. However, most studies use the dataset from the same source(s) for training and testing so that the ability of a trained DCNN to generalize to a different dataset is not well studied, as well as the strategy to address the issue of performance drop on a different dataset. In this study we investigated the performance of a well-trained DCNN model from a public dataset for thoracic OAR segmentation on a local dataset and explored the systematic differences between the datasets. We observed that a subtle shift of organs inside patient body due to the abdominal compression technique during image acquisition caused significantly worse performance on the local dataset. Furthermore, we developed an optimal strategy via incorporating different numbers of new cases from the local institution and using transfer learning to improve the accuracy and robustness of the trained DCNN model. We found that by adding as few as 10 cases from the local institution, the performance can reach the same level as in the original dataset. With transfer learning, the training time can be significantly shortened with slightly worse performance for heart segmentation.