Total knee arthroplasty (TKA) is effective for severe osteoarthritis and other related diseases. Accurate prosthesis prediction is a crucial factor for improving clinical outcomes and patient satisfaction after TKA. Current studies mainly focus on conventional manual template measurements, which are inconvenient and inefficient.
In this article, we utilize convolutional neural networks to analyze a multimodal patient data and design a system that helps doctors choose prostheses for TKA. To alleviate the problems of insufficient data and uneven distribution of labels, research on model structure, loss function and transfer learning is carried out. Algorithm optimization based on error correct output coding (ECOC) is implemented to further boost the performance.
The experimental results show the ECOC-based model reaches prediction accuracies of 88.23% and 86.27% for femoral components and tibial components, respectively.
The results verify that the ECOC-based model for prosthesis prediction in TKA is feasible and outperforms existing methods, which is of great significance for templating.