Boundary sensitive-net-based lumbar vertebra segmentation and spondylolisthesis measurement

Abstract Lumbar spine disorders represent a significant public health concern, with accurate diagnosis relying on vertebral segmentation and quantification. Traditional methods, such as Cobb angle measurement are constrained by two-dimensional projections, while cumulative segmentation and quantification errors limit automated CT analysis. To overcome these issues, this paper proposes a deep learning-based Boundary-Sensitive Network (BS-Net), integrating a Multi-Task Edge Processing (MEP) module and Contextual Bilateral Fusion (CBF) module to enhance vertebral edge feature extraction. The framework combines edge loss functions with morphological post-processing to achieve joint segmentation and quantification. Evaluations on 783 lumbar CT images from 379 patients and the public SPIDER MRI dataset demonstrate that BS-Net surpasses baseline models, achieving an MIoU of 96.56% and a Dice coefficient of 98.5%. Its spondylolisthesis quantification also shows strong agreement with manual assessment (ICC> 0.9). These results indicate that BS-Net provides an efficient and accurate solution for automated diagnosis of lumbar spondylolisthesis, with substantial clinical value.