Less is more: downsampling x-ray images improves pose estimation accuracy.

Area of Science:

• Biomechanics
• Medical Imaging
• Orthopedics

Background:

• Biplanar videoradiography (BVR) is a gold-standard for quantifying in vivo bone motion.
• The impact of X-ray image resolution on BVR pose estimation accuracy is not well understood.

Purpose of the Study:

• To investigate how downsampling X-ray images affects model-based pose estimation accuracy in BVR.
• To compare bicubic and nearest-neighbor interpolation methods for image downsampling.

Main Methods:

• High-speed BVR data from a participant with implanted tantalum beads were used.
• Images were downsampled from 2048 × 2048 to 512 × 512 using bicubic and nearest-neighbor interpolation.
• Pose estimation accuracy was evaluated by comparing downsampled images to full-resolution images across various bones and perturbation levels.

Main Results:

• Downsampling significantly reduced rotational and translational errors for both interpolation methods compared to full-resolution images.
• Bicubic interpolation showed slightly better pose accuracy for certain bones due to enhanced edge clarity.
• Full-resolution images resulted in more outliers and greater variability in pose estimates.

Conclusions:

• Downsampling X-ray images enhances pose estimation accuracy in BVR, even for complex joints like the ankle.
• Bicubic downsampling to 512 × 512 pixels is recommended as a best practice for BVR tracking of the ankle complex.
• This finding is applicable to both automated and manual BVR workflows.