Three-dimensional spine reconstruction from biplane radiographs using convolutional neural networks

Area of Science:

• Medical imaging
• Deep learning
• Spinal reconstruction

Background:

• Biplanar radiographs are commonly used for spinal imaging.
• Accurate 3D reconstruction of the spine is crucial for diagnosis and treatment planning.
• Current methods for 3D spine reconstruction from biplanar radiographs have limitations.

Purpose of the Study:

• To develop and evaluate a deep learning network for 3D spine reconstruction from biplanar radiographs.
• To improve the accuracy and reliability of 3D spinal models generated from 2D imaging data.

Main Methods:

• A deep learning network was designed to extract and reconstruct bone tissue features from biplanar radiographs.
• The network utilizes multiscale and similar feature extraction techniques.
• High-dimensional features were transformed into a 3D image domain through dimensionality reduction.
• Eight public datasets were used for training and testing the network.
• A combination of novel and classical evaluation metrics were employed.

Main Results:

• The deep learning method achieved a Hausdorff distance of 1.85 mm.
• Surface and volume overlap metrics demonstrated high reconstruction accuracy (0.2 mm and 0.9664, respectively).
• The method showed a minimal offset distance of 0.21 mm from the vertebral body centroid.

Conclusions:

• The developed deep learning network provides a reliable method for 3D spine reconstruction from biplanar radiographs.
• The approach demonstrates potential for enhanced accuracy in spinal imaging analysis.
• This technology could aid in more precise diagnosis and treatment planning for spinal conditions.