2D CT Vertebra Instance Segmentation for Computed-Assisted Spine Surgery

Abstract

Computer-Assisted Spine Surgery, in which the creation of 3D models plays an important role, provides great convenience to surgeons in preoperative surgical planning. For this purpose, computed tomography images are preferred because the bone structures are evident. Segmentation, the core process of creating a 3D model, is time-consuming when done manually. However, the segmentation has been shortened with recent developments in deep learning. Thus, we demonstrate an automated segmentation process with the Mask R-CNN model on the Detectron2 platform that enables rapid and correct prediction of vertebrae in 2D CT images. Index Terms―segmentation, deep learning, detectron2