Personalised pose estimation from single-plane moving fluoroscope images using deep convolutional neural networks.

Area of Science:

• Biomechanics
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate joint kinematics measurement is crucial for biomechanical research.
• X-ray based systems offer an alternative to skin-based methods, avoiding soft-tissue artifacts.
• Extracting 6D pose from X-ray images is traditionally time-consuming and resource-intensive.

Purpose of the Study:

• To develop an automated deep-learning model for precise 6D pose estimation of knee implants.
• To enhance the efficiency and reduce the cost of joint kinematics measurement.
• To provide personalized pose predictions for unseen subjects using pre-trained models.

Main Methods:

• Training a deep-learning model on over 106,000 annotated fluoroscopic images of knee implants.
• Utilizing a moving fluoroscope during activities of daily living for data collection.
• Employing pretraining with rendered implant geometries for personalized predictions.

Main Results:

• The model accurately estimated 6D poses for femoral and tibial components.
• Performance exceeded 0.75 mm (in-plane translation) and 2° (rotations) for 50% of samples.
• Robust performance was maintained even with occlusions and low contrast images (>90% samples).

Conclusions:

• The developed deep-learning approach enables fully automated and accurate 6D pose estimation of knee implants.
• This method significantly advances the field of biomechanical analysis and clinical applications.
• Personalized predictions and robustness in challenging imaging conditions are key advantages.