Toward Automated 3D Spine Reconstruction from Biplanar Radiographs Using CNN for Statistical Spine Model Fitting

Area of Science:

• Medical Imaging
• Spine Biomechanics
• Artificial Intelligence in Medicine

Background:

• Current 3D spine reconstruction from biplanar radiographs is labor-intensive and not clinically efficient.
• Existing semi-automated methods require significant user supervision and are time-consuming.

Purpose of the Study:

• To develop a novel, rapid, and automated 3D spine reconstruction technique.
• To enhance the clinical utility of 3D spine measurements.

Main Methods:

• Utilized convolutional neural networks (CNNs) to fit a statistical spine shape model to biplanar radiographs.
• CNNs automatically identify anatomical landmarks for spine model deformation.
• Employed a hierarchical, iterative process for landmark detection and model fitting.

Main Results:

• Achieved mean landmark location errors of 1.6-2.3 mm.
• Clinical parameter extraction showed mean errors of 2.8°-4.7° for spinal and 1°-2.1° for pelvic parameters.
• Automated measurements agreed with expert assessments 89% of the time, with reconstruction under one minute.

Conclusions:

• The proposed automated method offers a fast and accurate solution for 3D spine reconstruction.
• This advancement facilitates the integration of 3D spine measurements into routine clinical practice.
• The method shows high agreement with expert evaluations, supporting its clinical validity.