Abstract

Info

Title:: Spine CT image segmentation based on multi-scale boundary segmentation and hybrid attention mechanism

Author(s):: LIU Jing1; XU Hao1; CUI Xinxin1; TIAN Zhenyu1; YANG Jianlan2; 1. School of Information Engineering, Gansu University of Chinese Medicine, Lanzhou 730000, China 2. Quanzhou Orthopedic-Traumatological Hospital, Quanzhou 362019, China

Keywords:: Keywords: spine segmentation 3D U-Net vertebral boundary segmentation hybrid attention mechanism

Abstract:: The early diagnosis of spinal diseases is mainly screened and initially diagnosed through computed tomography (CT). In view of the complex structure of vertebral bones and low segmentation accuracy in spinal CT images, a spinal CT image segmentation network based on 3D U-Net framework is proposed. The network which integrates squeeze-and-excitation residual module, vertebral boundary segmentation model, and improved hybrid channel-spatial attention mechanism is trained and tested on VerSe 19, VerSe 20, and CTSpine1K spinal datasets. Multiple experiments indicate that the model can effectively improve segmentation accuracy and efficiency while demonstrating good generalization and robustness. Compared with other advanced network models, the proposed network achieves higher segmentation accuracy in terms of Dice similarity coefficient, Hausdorff distance, and average symmetric surface distance. The proposed model exhibits superior segmentation performance among the existing spinal segmentation networks, providing radiologists with valuable clinical information.