An Effective CNN Approach for Vertebrae Segmentation from 3D CT Images

Spine surgery is risky when treating spinal diseases or injuries. In order to decrease the damage of tissue and accelerate the recovery, minimally invasive surgery is strongly desired. Because of this, the preoperative planning and intraoperative guidance become very important. A 3D vertebrae model of the surgical site will be built for these purposes from the computer tomography (CT) images. The segmentation methods of vertebrae from the CT images for modeling are widely concerned. The conventional methods were satisfactory but they usually required prior knowledge of the vertebrae. Recently, the deep learning convolutional neural network (CNN) showed an outstanding performance on the end-to-end semantic image segmentation. Benefit from this, we propose an effective and simple convolution neural network based approach for vertebrae segmentation. Different from training and testing on large 3D data, the axial view images are used effectively in this study. A sequence of simple image processing steps are applied for the intervertebral plane detection. Five CT cases are used on this study and the segmentation results are highly accurate with average dice similarity coefficient over 0.95.