In this paper, we propose a fully automatic method for vertebra segmentation in the CT volume data. The method constructs 3D fences that separate adjacent vertebrae from valley-emphasized Gaussian images. Initial curves for the 3D fences are extracted from intervertebral discs, detected with anatomical characteristics, then optimized using a deformable model. A minimum cost path finding method corrects any erroneous curves trapped into a local minimum. Final volume is labeled with help of the 3D fences by a fence-limited region growing method. This method has been applied to 50-patient data sets and has proved to be very successful.