INSTANCE-LEVEL LOCALIZATION, SEGMENTATION, AND IDENTIFICATION OF VERTEBRAE IN 3D CT IMAGES

Abstract

The spine's complex anatomy and essential functions necessitate precise vertebral analysis for numerous clinical applications, including pre-surgical planning, accurate pathological diagnosis, fracture detection, and comprehensive postoperative assessments. Computed tomography (CT) imaging, recognized for its high-resolution and three-dimensional visualization, remains a fundamental tool in spinal evaluations. Despite its capabilities, automating vertebral localization, segmentation, and identification in CT scans presents significant challenges due to the variability in scan coverage, patient anatomy, and vertebral morphology. These challenges are further exacerbated by substantial intra-class variability among patients and inter-class similarities between adjacent vertebrae, especially within transitional regions such as the cervicothoracic (C7-T1) and thoracolumbar (T12-L1) junctions. Such factors can lead to misclassifications and compromise diagnostic accuracy. To address these issues, we propose a two-stage multi-task pipeline that combines dual-label supervised contrastive learning for pretraining with sequence maximum likelihood optimization. This approach is tailored to enhance the differentiation of vertebrae by leveraging both individual vertebral features and contextual anatomical information, thereby improving identification accuracy in critical areas of the spine. Extensive validation on the VerSe'19 and VerSe'20 datasets demonstrates that our method outperforms current state-of-the-art techniques, delivering substantial improvements in vertebrae identification accuracy. These advancements not only optimize vertebral analysis but also hold the potential to significantly enhance clinical outcomes in a wide range of spinal diagnostic and therapeutic procedures.