Fully automated 3D segmentation and separation on multiple cervical vertebrae of CT images using 2D deep convolutional neural net: validation with intra- and extra-dataset

Heejung Hyun, Younghwa Byeon, Yongwon Cho, Namkug Kim, Seong Yi, Sung-Uk Kuh, Jin S. Yeom

Apr 11, 2018 (modified: May 16, 2018) MIDL 2018 Abstract Submission readers: everyone
  • Abstract: In this study, we developed a fully-automated cervical vertebrae segmentation and separation model based on 2D deep convolutional neural networks (CNN), and evaluated it with intra- and extra-datasets. The main idea of this research was to label superior and inferior vertebrae differently, to train the 2D U-net model for semantic segmentation, and post-process the predicted results for 3D vertebra separation. Spine CT scans of 17 patients with cervical spine injury and 19 healthy subjects, were collected from two independent centers. The model was tested with various evaluation metrics including Dice similarity coefficeint (DSC), Jaccard similarity coefficient (JSC), mean surface distance (MSD) and Hausdorff surface distance (HSD). Despite the data was from subjects with various clinical and demographic information and was acquired using different CT scan protocols, the proposed model has shown robust and good accuracy.
  • Keywords: cervical vertebrae, convolutional neural network (CNN), deep learning, spine segmentation, spine CT, U-Net
  • Author affiliation: Asan Medical Center, Severance Hospital, Gangnam Severance Spine Hospital, Seoul National University Bundang Hospital
0 Replies

Loading