A graph-based spine segmentation method: Combining target detection with unsupervised segmentation

Spine image segmentation is important for the diagnosis and treatment of spinal diseases. However automatic segment vertebrae and intervertebral discs from spine images without segmentation labels is a challenging. In this paper, we propose an end-to-end spine image segmentation framework to achieve automated spine image segmentation. The framework consists of an initialization stage, a coarse segmentation stage and a fine segmentation stage. The initialization stage is a trained regions of interest detector. The coarse segmentation stage is a deep autoencoder clustering network. In this stage, the reconstruction loss and clustering loss of is used to achieve unsupervised coarse segmentation. In addition, a fixed number of channels strategy is also employed to greatly reduce the number of model parameters while ensuring the segmentation performance. In the fine segmentation stage, the image segmentation is reinterpreted from the perspective of the graph structure. The edge pixels from the coarse segmentation are used to construct graphs. The features of the nodes and the features of the edges between nodes are fully considered by the graph attention mechanism to achieve unsupervised fine segmentation. Experiments on two spine image segmentation datasets and one brain tumor image segmentation dataset show that our method has superior performance and generalization ability.