Validation of a machine learning technique for segmentation and pose estimation in single plane fluoroscopy

Area of Science:

• Orthopedic Surgery
• Biomedical Engineering
• Medical Imaging

Background:

• Total knee replacement (TKR) kinematics are crucial for evaluating surgical success.
• Current kinematic measurement methods are time-consuming and labor-intensive, limiting clinical application.

Purpose of the Study:

• To validate a novel, automated single-plane fluoroscopic technique for measuring TKR kinematics.
• To compare the accuracy and reproducibility of the automated method against biplane radiostereometric analysis (RSA).

Main Methods:

• Utilized machine learning with two neural networks for automated segmentation and pose estimation of TKR components.
• Simulated single-plane fluoroscopy using lateral images from biplane RSA data (n=113 image pairs from 24 knees).
• Compared automated single-plane kinematics with manual biplane RSA using root-mean-square error and Bland-Altman analysis.

Main Results:

• Achieved root-mean-square errors of 0.8 mm (AP), 0.5 mm (SI), 2.6 mm (ML) for translations and 1.0° (FE), 1.2° (AA), 1.7° (IE) for rotations.
• Demonstrated high reproducibility with submillimeter errors for in-plane translations and <2° for all rotations between observers.
• The automated technique showed strong agreement with the gold standard biplane RSA.

Conclusions:

• The automated single-plane fluoroscopic technique provides accurate and reproducible TKR kinematic measurements.
• This method significantly simplifies and speeds up TKR kinematic analysis, making it more accessible for clinical practice and research.
• The findings support the integration of this automated technique into routine clinical workflows for improved TKR assessment.