In the spinal computed tomography (CT) image segmentation problem, owing to the low contrast between the spine and tissues, and the influence of noise, the traditional segmentation algorithms have problems such as poor segmentation accuracy and low degree of automation. Aiming at solving the above-mentioned problems, a method of locating the spine through AttentionNet and then using improved DenseUnet to perform spinal CT segmentation is proposed herein. First, preprocessing operations such as cropping, resampling, and normalization of gray values are performed on all spinal CT sample data; the samples are trained using AttentionNet to obtain Attention maps with position information. Second, the traditional DenseUnet is improved, and each Dense block adds the Shuffle operation to increase the network robustness. After each Dense block, a 1×1 convolution is added to reduce the number of channels and network parameters. Third, the training samples are pretrained using the improved DenseUnet to obtain the prediction maps with prior information. Finally, the Attention map, prediction map, and the original images are fused into three-channel training samples as the input, and the improved DenseUnet is used to train segmentation model and is verified against the test set. Consequently, the spinal CT automatic segmentation is realized. The experimental results show that the segmentation accuracy of the proposed method is better than that of the traditional DenseUnet, and the proposed method is an effective automatic segmentation method for spinal CT.