The paper presents a two-stage deep learning framework for vertebral and intervertebral disc segmentation on T2-weighted MR images in spine injury analysis. Most existing CNN-based methods segment vertebrae and discs independently, thereby losing their anatomical relationships. Modeling the spine as a graph with nodes and an adjacency matrix allows these relationships to be exploited. The first stage uses a 3D graph convolutional network (GCSN) for coarse multi-class segmentation. The second stage applies a 2D ResNet network to improve boundary resolution. They tested the model on data from 218 subjects with an average cube similarity coefficient (DSC) of 87.32% on 10 vertebrae, 87.78% on 9 discs and 87.49% on 19 structures. Results show improved consistency and accuracy due to anatomical dependencies. The system is used for automated analysis of the spine for diagnosis and treatment planning of disorders.