Vertebrae segmentation via stacked sparse autoencoder from computed tomography images

An accurate vertebrae segmentation in the spine is an essential pre-requisite in many applications of image-based spine assessment, surgical planning, clinical diagnostic treatment, and biomechanical modeling. In this paper, we present the stacked sparse autoencoder (SSAE) model for the segmentation of vertebrae from CT images. After the preprocessing step, we extracted overlapped patches from the vertebrae CT images as the inputs of our proposed model. The SSAE model was trained in an unsupervised way to learn high-level features from the input pixels of the unlabeled images patch. To improve the discriminability of the learned features, we further refined the feature representation in a supervised fashion and fine-tuned the whole model by using the feedforward neural network parameters for classifying the overlapped patches. We then validated our model on a publicly available MICCAI CSI2014 dataset and found that our model outperforms the other state-of-the-art methods.