A deep learning framework for segmentation and pose estimation of pedicle screw implants based on C-arm fluoroscopy

Area of Science:

• Spinal surgery
• Medical imaging
• Artificial intelligence in medicine

Background:

• Pedicle screw fixation is crucial for spinal stability but has high reoperation rates.
• Current intraoperative verification relies on visual assessment of radiographic images, which has limited accuracy.
• There is a need for automated systems to improve pedicle screw insertion assessment.

Purpose of the Study:

• To develop an accurate and automated system for intraoperative pedicle screw assessment.
• To utilize deep learning for automatic segmentation and pose estimation of pedicle screws.
• To enhance the precision of pedicle screw placement and reduce revision surgeries.

Main Methods:

• A convolutional neural network was employed for segmenting pedicle screw X-ray projections into screw head, shaft, and background.
• Biplanar X-ray spatial configuration knowledge was used to establish correspondence between projections.
• A 6-degree-of-freedom pose estimation was performed for each segmented pedicle screw.

Main Results:

• The machine learning framework achieved 93% accuracy in segmenting screw shafts on synthetic X-rays and 83% on clinical X-rays.
• Pose estimation accuracy on clinically realistic X-rays was demonstrated.
• The system provides a robust method for evaluating pedicle screw placement.

Conclusions:

• The proposed system offers an accurate, fully automatic framework for pedicle screw segmentation and pose assessment.
• This technology can aid in developing intraoperative pedicle screw insertion assessment protocols.
• The system integrates seamlessly with existing surgical workflows, minimizing disruption.