Annotations for medical images are very hard to acquire as it requires specific domain knowledge. Therefore, performance of deep learning algorithms on medical image processing is largely hindered by the scarcity of large-scale labeled data. To address this challenge, we propose a semi-supervised learning method for lesion detection from CT images which exploits a key characteristic of the volumetric medical data, i.e. adjacent slices in the axial axis resemble each other, or say they bear some kind of continuity. Specifically, by exploiting such a prior, a semi-supervised scheme is adopted to propagate bounding box annotations to adjacent CT slices to obtain more training data with fewer false positives and more true positives. Furthermore, considering that the NIH DeepLesion dataset has many missing labels, we develop a missing ground truth mining process by considering the continuity (or appearance-consistency) of multi-slice axial CT images. Experimental results on the NIH DeepLesion dataset demonstrate the effectiveness our methods for both semi-supervised label propagation and missing label mining.