Comprehensive Summary
This study evaluates deep learning (DL) models for segmentation of computed tomography (CT) images to detect tibial implant loosening. Instead of the current semi-automatic method of CT image segmentation, the researchers develop a DL model that does not require human intervention. Three different datasets (cadaver, patient, and reproducibility) were used to develop and evaluate the model. Six nnU-Net models (Cortex 2D, Cortex 3D, Full 2D, Full 3D, Multi-Class 2D, Multi-Class 3D) were trained. The best performing model, Cortex 3D, was selected as the final model as it had high segmentation performance of 95.56% Dice similarity coefficient (DSC) and 0.58 mm 95th percentile Hausdorff Distance (HD95) metric. The performance of the model demonstrate potential of the automatic segmentation to replace the current semi-automatic method.
Outcomes and Implications
Implant loosening is a big issue for patients with total knee arthroplasty (TKA), being the cause of 30% of the revision surgery done with in 10 years. The current noninvasive method shows indirect signs of loosening, opening up the possibility of misdiagnosis. It is only possible to measure the migration of implant over time where the current approach involves processing valgus and varus loaded CT scans, which requires manual correction during segmentation. Due to the human resource required, there is a need for a fully automated method. The researchers propose a DL model in this study, which has the potential to replace the semi-automatic method.