Vertebra CT image segmentation is the key to the visualization of vertebra 3D reconstruction. Aiming at the problems of blurred vertebra edge, complex structure and changeable shape in vertebra CT images, a dual-decoder network is proposed based on deep learning method. The network adds a parallel decoding branch with the same structure on the basis of the U-Net structure of the encoding and decoding network, and the two decoding branches can extract image features complementary. Moreover, a dual feature fusion module is added between encoding and decoding to solve the problem of semantic information loss caused by the network downsampling and upsampling. At the same time, the original convolution module is replaced by the densely connected hybrid convolution module to improve the network's ability to extract multi-scale features. In addition, an efficient attention module is added to make the network focus on learning regions of interest in space and suppress irrelevant information in channels. Tested on the CSI2014 public dataset, the Dice coefficient reaches 0.970, the Jaccard coefficient reaches 0.945, and the Recall rate reaches 0.962. The experimental results show that the network can improve the accuracy of vertebra segmentation, has better generalization ability, and can meet the needs of clinical vertebra CT image segmentation.