【Abstract】 In the field of medical imaging, accurate and rapid spinal instance segmentation is essential for early diagnosis and treatment planning. Aiming at addressing the problems of low efficiency and poor performance of previous spinal instance segmentation methods, an innovative two-stage 3D spinal instance segmentation method based on local window iteration is proposed in this paper. In the first stage, semantic segmentation of the whole spine is carried out by nn-UNet model to roughly identify the positions and shapes of all relevant structures, providing necessary contextual information for the instance segmentation stage. In the second stage, an innovative case segmentation method based on local window iteration is used to partition the window through the centroid of the spinal unit obtained based on the semantic segmentation results, and all window iteration instances are segmented to accurately segment each spinal unit. Finally, the complete spinal instance segmentation results are obtained through post-processing reconstruction. In this paper, the instance segmentation task is innovatively defined as the segmentation center and the upper and lower adjacent spinal units. By simplifying the complex instance segmentation task, the training difficulty of the model is reduced effectively and the segmentation accuracy is improved. The experimental results show that the proposed method outperforms previous spinal segmentation methods on SPIDER and CSI2014 data sets, demonstrating its superiority and potential in dealing with complex spinal structures.