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

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Spinal imaging

Background:

• Accurate vertebral landmark localization in computed tomography (CT) is essential for diagnosing spinal conditions, planning surgeries, and assessing outcomes.
• Challenges persist in precisely locating vertebrae within high-resolution 3D CT datasets due to large anatomical spans and similar vertebral morphologies.

Purpose of the Study:

• To introduce a novel deep learning framework for robust and precise vertebral landmark localization in 3D CT volumes.
• To overcome the limitations of existing methods in handling complex spinal anatomy.

Main Methods:

• A two-stage deep learning approach was developed, starting with an improved squeeze and excitation V-Net (ISE-VNet) for initial spinal column segmentation.
• The second stage employed a 3D generalized differential spatial-to-numerical transform (DSNT) module for accurate individual vertebral localization.

Main Results:

• The proposed framework achieved a significant improvement in identification rates, increasing from 81.15% to 96.32%.
• Localization errors were substantially reduced from 7.6 mm to 2.1 mm, surpassing current state-of-the-art methods.

Conclusions:

• This deep learning framework offers a reliable solution for clinical vertebral landmark localization in 3D medical imaging.
• The method facilitates efficient, focused analysis of vertebral regions, proving valuable for both research and clinical applications.