Enhanced vertebrae localization in CT volumes: a two-stage deep learning framework.

BACKGROUND: Vertebral landmark localization in computed tomography (CT) volumes is crucial for spinal pathological diagnosis, postoperative assessment and surgical planning. However, vertebral localization within high-resolution three-dimensional CT volumes still poses prominent challenges. METHODS: This study introduces a novel two-stage deep learning framework that effectively addresses challenges such as large spatial spans and high morphological similarity of vertebrae. The first stage employs the improved squeeze and excitation V-Net (ISE-VNet) for coarse segmentation of the spinal column. The second stage utilizes a 3D spatial-generalized differential spatial-to-numerical transform (DSNT) module for precise localization of individual vertebrae. RESULTS: Our framework significantly improves identification rates from 81.15% to 96.32% and reduces localization errors from 7.6 mm to 2.1 mm, outperforming state-of-the-art methods. CONCLUSION: This approach provides a robust solution for clinical vertebral landmark localization in 3D medical images. This design effectively enables efficient and focused analysis of vertebral local regions and it holds significant value in both scientific research and clinical practice.