Automatic Localization and Segmentation of Vertebral Bodies in 3D CT Volumes with Deep Learning

Automatic localization and segmentation of vertebral bodies in CT volumes bears many clinical utilities, such as shape analysis. Variation in the vertebra appearance, unknown field-of-views, and pathologies impose several challenges for these tasks. Most previous studies targeted the whole vertebra and their algorithms, though were of high accuracy, made high demand on hardware and took longer than feasible in daily clinical practice. We developed a two-step algorithm to localize and segment just vertebral bodies by taking the advantage of the intensity pattern along the front spinal region, as well as GPU accelerations using convolutional neural networks. First, we designed a 2D U-net variants to extract front spinal region, based on which the centroids of vertebra were localized using M-method and 3D region of interests were generated for each vertebra. Second, we developed a 3D U-net with inception module using dilated convolution to segment vertebral bodies in the 3D ROIs. We trained our two U-nets on 61 annotated CT volumes. Tested on three unseen CTs, our methods achieved an identification rate of 92% and detection error 0.74 mm and Dice coefficient of 0.8 for the 3D segmentation using less than 10 seconds per case.