Automatic Segmentation of Human Spine with Deep Neural Network

Considering that a CT scan produces cross-sectional images of specific areas of a scanned object, medical images of the spine are much more complex than those of other organs. In this study, the entire task of CT segmentation is viewed as a binary classification problem. We modify two deep learning networks—U-Net and SegNet—that are often used in the field of semantic segmentation. We made following modifications. Firstly, considering the size of CT images, we further reduce the number of convolutional layers. Then we use an element-wise method instead of concatenation in U-Net. Lastly, we select a new loss function as an evaluation criterion. According to the experimental results, we conclude that U-Net is not applicable when the training set is large, in which case we cannot prevent overfitting. However, SegNet performs better than U-Net in CT images segmentation.