Performance of local optimization in single-plane fluoroscopic analysis for total knee arthroplasty

Area of Science:

• Biomedical Engineering
• Medical Imaging Analysis
• Orthopedic Surgery

Background:

• Fluoroscopy-derived joint kinematics is crucial for knee prosthesis evaluation.
• Accurate 3D prosthesis pose estimation from 2D fluoroscopic images is essential but currently labor-intensive and operator-dependent due to extensive user interaction.

Purpose of the Study:

• To review and assess the performance of five optimization methods for 3D pose estimation of knee prostheses.
• To compare derivative-free and gradient-based optimizers, including implementations with numerical and analytical gradients.

Main Methods:

• Evaluated two derivative-free optimizers (DHSAnn, IIPM) and three gradient-based optimizers (LevMar, DoNLP2, IpOpt).
• Compared numerical and analytical gradient implementations for gradient-based methods.
• Assessed performance on both phantom and clinical fluoroscopic data.

Main Results:

• All tested methods achieved >85% success rate for 3D pose estimation within 1mm and 1° on phantom data.
• IpOpt demonstrated the highest success rate (97%) on phantom data.
• Clinical data showed >85% success for in-plane positions but lower rates for rotations; analytical gradients accelerated computation 3-4x without impacting success rates.

Conclusions:

• Approximately 85% of clinical fluoroscopy frames can be analyzed automatically, with only 15% requiring manual supervision.
• The high success rate on phantom data suggests that further reductions in manual supervision may be achievable with optimized methods like IpOpt.