Vertebrae localization, segmentation and identification using a graph optimization and an anatomic consistency cycle.

Area of Science:

• Medical Imaging
• Computer Vision
• Biomedical Engineering

Background:

• Accurate vertebrae localization, segmentation, and identification in CT images are crucial for clinical applications.
• Deep learning methods show promise but struggle with transitional and pathological vertebrae due to limited training data.
• Non-learning methods utilize prior knowledge but may lack generalizability.

Purpose of the Study:

• To develop an improved method for vertebrae identification in CT images by combining deep learning and statistical priors.
• To enhance the handling of transitional and pathological vertebrae, which are often underrepresented in datasets.
• To achieve state-of-the-art performance on vertebrae identification benchmarks.

Main Methods:

• An iterative deep learning approach for recurrent vertebrae localization, segmentation, and identification.
• Integration of statistical priors to enforce anatomical consistency throughout the spine.
• Utilizing a graphical model to aggregate local predictions for robust transitional vertebrae identification.

Main Results:

• Achieved state-of-the-art results on the VerSe20 challenge benchmark.
• Outperformed existing methods in identifying transitional vertebrae.
• Demonstrated strong generalization capabilities on the VerSe19 challenge benchmark.
• Successfully detected and reported anatomically inconsistent spine regions.

Conclusions:

• The combined deep learning and statistical prior approach significantly improves vertebrae identification accuracy in CT images.
• This method offers enhanced performance for challenging cases like transitional vertebrae.
• The approach provides a robust and generalizable solution for spinal analysis and can identify anatomical inconsistencies.