Vertebral body segmentation in CT: An open dataset, deep-learning models and comparison to existing models

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence

Background:

• Vertebral bodies serve as crucial anatomical landmarks for quantitative measurements in medical imaging.
• Accurate localization of vertebral bodies is essential for applications like body composition analysis.
• Existing methods for vertebral body segmentation may have limitations in accuracy and accessibility.

Purpose of the Study:

• To develop and provide open-access deep-learning models for vertebral body segmentation.
• To create open-access vertebral body labels for research and development.
• To compare the performance of novel segmentation models against existing solutions for identifying the third lumbar vertebra (L3).

Main Methods:

• Thoracic and lumbar vertebral body labels were generated from 1460 CT scans across two public datasets.
• Two residual-encoder nnU-Net models and a two-step pipeline were trained and evaluated for segmentation performance.
• Performance was assessed on an independent dataset using localization error and center hit rate metrics, comparing against established models.

Main Results:

• The developed models achieved high Dice scores (0.962 and 0.939) for L3 vertebral body segmentation.
• Segmentation models demonstrated high center hit rates and lower absolute localization error compared to the TotalSegmentator pipeline.
• The proposed models offer reliable identification of L3 with comparable or superior performance to existing methods.

Conclusions:

• The novel deep-learning models provide accurate segmentation of thoracic and lumbar vertebral bodies.
• These models reliably identify the third lumbar vertebra (L3), crucial for anatomical localization.
• Open-access labels, model weights, and a body composition analysis pipeline are now available to the research community.