Open vertebrae CT dataset and deep learning models benchmarked against existing tools

Open neural network models for vertebral body segmentation in computed tomography (CT) achieve Dice scores up to 0.962 (95% CI 0.941–0.978) for L3 and outperform existing pipelines in center localization error, externally validated in 300 oncologic cases. Labels and weights are…

• VertebralBodiesL and VertebralBodiesStepwise showed lower absolute center localization error than TotalSegmentator pipeline.
• Signed center localization error did not differ significantly across models.
• The dataset, model weights, and body composition analysis pipeline are openly available.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.