LSRC: A Long-Short Range Context-Fusing Framework for Automatic 3D Vertebra Localization

Abstract

Automatic localization and identification of vertebrae in computed tomography (CT) images is a challenging task, due to the specific spine structure, complex pathological conditions, and limited field-of-view in 3D CT images. The local and long-range contextual information is especially useful for solving this problem. To explore both the local and long-range contextual information of vertebrae, in this paper, we propose a new framework called Long-Short Range Context-fusing framework (LSRC), combining a 3D local semantic network and a 2D long-range contextual network. The 3D local semantic network, using 3D CT images, produces 3D heat maps corresponding to the locations of all vertebrae. The 3D heat maps and CT images are respectively projected onto the sagittal plane and coronal plane, and fed to the 2D long-range contextual network, in which a convolutional encoder-decoder module integrates the long-range contextual information of these two views. Two refined heat maps in the sagittal and coronal planes are generated by a globally-refining module, adjusting vertebra locations using the global location information in an attention manner. Experiments on a public dataset of 302 3D spine CT scans with various pathological conditions show that our new framework outperforms state-of-the-art methods.