Development and model form assessment of an automatic subject-specific vertebra reconstruction method

Area of Science:

• Biomedical Engineering
• Medical Imaging
• Computational Anatomy

Background:

• Current 3D spine models rely on generic databases or laborious manual data.
• Existing methods are time-consuming and may lack subject-specific accuracy.
• This limits their application in advanced surgical planning and training.

Purpose of the Study:

• To develop and validate a rapid, accurate workflow for subject-specific vertebra reconstruction.
• To quantify the geometric accuracy and model form errors of the reconstructed 3D models.
• To assess the biomechanical implications of these models using finite element analysis.

Main Methods:

• Customized four neural networks for vertebra segmentation.
• Reconstructed 3D CAD models from CT scans of an excised human lumbar vertebra.
• Employed a reverse engineering approach for gold-standard geometry.
• Utilized 3D evaluation metrics (Dice, IoU, Hausdorff distance) and finite element analysis (FEA).

Main Results:

• Achieved a high Dice score of 94.20% for automatic segmentation.
• Reconstructed models showed excellent accuracy with a 3D Dice index of 92.80% and 3D IoU of 86.56%.
• FEA demonstrated biomechanical accuracy with a closest percentage error of 4.2710% compared to the gold standard.

Conclusions:

• An automated workflow for subject-specific vertebra reconstruction was successfully developed.
• Quantified geometric and FEA errors provide crucial insights for clinical application.
• This method offers a more accurate basis for developing and improving spine treatments.