Automation in tibial implant loosening detection using deep-learning segmentation

Area of Science:

• Biomedical Engineering
• Medical Imaging
• Artificial Intelligence

Background:

• Recurrent complaints after total knee arthroplasty (TKA) may indicate aseptic implant loosening.
• Current imaging methods struggle to quantify TKA component looseness.
• A validated workflow exists for quantifying tibial component displacement using loaded CT scans, but requires manual segmentation.

Purpose of the Study:

• To investigate the feasibility of fully automating the segmentation step in TKA component displacement analysis.
• To evaluate if deep learning (DL) models can replace semi-automatic segmentation without affecting outcome accuracy.
• To maintain the capability of distinguishing between fixed and loose implants.

Main Methods:

• Developed and evaluated various deep learning (DL) models for fully automatic segmentation of tibial components and bone from CT scans.
• Utilized three datasets for model development and evaluation, including cadaveric and patient CT data.
• Integrated the best-performing DL model into the existing workflow, replacing the semi-automatic segmentation step.

Main Results:

• The DL-based approach demonstrated statistically significant differences between fixed and loose implant samples in both cadaveric and patient datasets.
• Methodological errors were not significantly different between the automated and semi-automatic approaches.
• Both the proposed and current approaches showed excellent reliability across multiple operators and datasets.

Conclusions:

• Full automation of knee implant displacement assessment is achievable using DL-based segmentation.
• The automated approach maintains the accuracy and reliability of the existing workflow.
• This automation can simplify the analysis of tibial component loosening after TKA.