Labeled data is scarce in clinical practice, and labeling 3D medical data is time-consuming. The study aims to develop a deep learning network with a few labeled data and investigate its segmentation performance of lumbosacral structures on thin-layer computed tomography (CT). In this work, semi-cGAN and fewshot-GAN were developed for automatic segmentation of nerve, bone, and disc, compared with 3D U-Net. For evaluation, dice score and average symmetric surface distance are used to assess the segmentation performance of lumbosacral structures. Another dataset from SpineWeb was also included to test the generalization ability of the two trained networks. Research results show that the segmentation performance of semi-cGAN and fewshot-GAN is slightly superior to 3D U-Net for automatic segmenting lumbosacral structures on thin-layer CT with fewer labeled data.