X23D-Intraoperative 3D Lumbar Spine Shape Reconstruction Based on Sparse Multi-View X-ray Data

Area of Science:

• Medical Imaging
• Computer Vision
• Spine Surgery

Background:

• Intraoperative 2D X-rays are crucial for surgical guidance but struggle with 3D anatomical assessment, particularly for complex structures like the spine.
• Accurate 3D visualization of the spine from planar fluoroscopic images presents a significant challenge for surgeons.

Purpose of the Study:

• To develop a novel deep learning method for intraoperative 3D lumbar vertebrae shape estimation from sparse, multi-view X-ray data.
• To improve the accuracy of 3D reconstructions by integrating X-ray calibration parameters into a learned multi-view stereo approach.

Main Methods:

• A deep learning-based multi-view stereo approach was developed to estimate 3D lumbar spine shape from X-ray images.
• The method incorporated X-ray calibration parameters into the neural network to leverage prior knowledge of spinal anatomy while maintaining patient specificity.
• The model was trained and validated on 17,420 fluoroscopy images derived from the CTSpine1K dataset.

Main Results:

• The method achieved an 88% average F1 score and a 71% surface score on unseen data.
• Incorporating calibration parameters improved the surface score by 22% compared to a state-of-the-art counterpart method.
• High-quality and accurate 3D reconstructions were demonstrated, outperforming existing techniques.

Conclusions:

• The proposed deep learning method enables accurate intraoperative 3D lumbar spine reconstruction from multi-view X-rays.
• This advancement offers new possibilities for surgical navigation and decision-making, especially in workflows lacking standard 3D imaging.
• The technique enhances patient specificity and accuracy, addressing limitations of current 2D-based assessments in spinal surgery.