Detecting loosening of total knee arthroplasty (TKA) implants is a challenging process, even for experienced clinicians, and may not always be available, potentially delaying diagnosis. The purpose of this study was to investigate whether the loosening of TKA implant could be detected on plain radiograph using deep convolution neural network (CNN).
An analysis was conducted on 100 patients who underwent revision TKA due to loosening at one institution from 2012 to 2020. 100 patients who underwent primary TKA without loosening were extracted through propensity score matching for age, gender, body mass index, and operation side. We implemented training acceleration in a pre-trained CNN model called VGG19 through a transfer learning method that gradually changes the freezing layer. Training and testing were conducted in a ratio of 8:2.
In the first model that trained only the last layer of the classification layer, the accuracy in diagnosing loosening of TKA implant was 87.5% and the accuracy for no-loosening was 56.3%. However, as the trained learning layer was gradually increased to the feature learning part, the accuracy in diagnosing loosening increased to 100%, and the accuracy for the non-loosening increased to 95%.
CNN algorithm through transfer learning shows high accuracy in detecting loosening of TKA implant through plain radiograph. These results can be expected to be utilized as an auxiliary tool in the decision making process in diagnosing loosening of TKA implant by orthopedic surgeons.