Accurate segmentation of cardiac medical images is essential for diagnosing and analyzing cardiovascular diseases. However, the scarcity of labeled cardiac Computed Tomography (CT) images makes it challenging to train neural networks using a fully supervised approach. In this paper, we propose a novel semi-supervised cardiac CT segmentation method, called SCAN. The key idea of SCAN is to fuse the pseudo labels from two deep neural networks to leverage unlabeled cardiac CT images. SCAN applies a two-stage framework. In the first stage, SCAN leverages a data perturbation strategy to enrich CT images with pseudo labels. In the second stage, SCAN employs a cross-supervision strategy to fuse the pseudo labels from two neural networks, using the outputs from each network as pseudo labels to train the other one. Finally, we evaluate SCAN on three public datasets of cardiac CT images. Experiments demonstrate that SCAN achieves the best performance and outperforms state-of-the-art semi-supervised segmentation methods by a large margin. It shows great potential for cardiac CT segmentation.