Background: Optimal implant position and alignment remains a controversial, yet critical topic in primary total knee arthroplasty (TKA). Future study of ideal implant position will require the ability to efficiently measure component positions at scale. Previous algorithms have limited accuracy, do not allow for human oversight and correction in deployment, and require extensive training time and dataset. Therefore, the purpose of this study was to develop and validate a machine learning model that can accurately automate, with surgeon directed adjustment, implant position annotation.
Methods: A retrospective series of 295 primary TKAs was identified. The femoral-tibial angle (FTA), distal femoral angle (dFA), and proximal tibial angle (pTA) were manually annotated from the immediate short leg post-op radiograph. We then trained a neural network to predict each annotated landmark using a novel label augmentation procedure of dilation, reweighting, and scheduled erosion steps. The model was compared against diverse models and accuracy was assessed using a validation set of 43 patients and test set of 79 patients.
Results: Our proposed model significantly improves accuracy compared to baseline training models across all measures in ten out of eleven cases (p < 1e-22 for each measure). The mean absolute error (difference from manual annotation) was 0.65° for FTA, 1.62° for dFA, and 1.44° for pTA.
Conclusion: Utilizing a novel algorithm, trained on a limited dataset, the accuracy of component position was approximately 1.2°. Additionally, the model outputs adjustable points from which the angles are calculated, allowing for clinician oversight and interpretable diagnostics for failure cases.
Keywords: Image registry; Machine learning; Primary TKA.
Copyright © 2024. Published by Elsevier Ltd.