Convolutional Neural Networks (CNNs) and Transformers have recently demonstrated promising performance in a multitude of supervised medical image segmentation tasks. However, the success of these methods heavily depends on massive annotated data for training, which is time-consuming and laborious to acquire, especially for pixel-wise annotation that requires medical expertise and clinical experience. To address this issue, we present a novel framework called Collaborative Transformer-CNN Learning (CTCL) for semi-supervised medical image segmentation. Specifically, Our CTCL combines cross teaching with hard pseudo labels (CTH) and mutual learning with soft pseudo labels (MLS) to simultaneously train the Transformer-CNN models, thus making full use of the unlabeled data for the supervision of the segmentation task while reducing the uncertainty and noises of pseudo labels. Furthermore, considering the large structural differences between the Transformer and CNN models, we utilize consistency regularization (CR) to make the predictions consistent and determined via minimizing classifier determinacy discrepancy. Extensive experiments on a popular medical image benchmark demonstrate that our method yields superior results compared with existing state-of-the-art semi-supervised learning methods.