Abstract
Background
Implant loosening is a common cause of a poor outcome and pain after total knee arthroplasty (TKA). Despite the increase in use of expensive techniques like arthrography, the detection of prosthetic loosening is often unclear pre-operatively, leading to diagnostic uncertainty and extensive workup. The objective of this study was to evaluate the ability of a machine learning (ML) algorithm to diagnose prosthetic loosening from pre-operative radiographs, and to observe what model inputs improve the performance of the model.
Methods
754 patients underwent a first-time revision of a total joint at our institution from 2012–2018. Pre-operative X-Rays (XR) were collected for each patient. AP and lateral X-Rays, in addition to demographic and comorbidity information, were collected for each patient. Each patient was determined to have either loose or fixed prosthetics based on a manual abstraction of the written findings in their operative report, which is considered the gold standard of diagnosing prosthetic loosening. We trained a series of deep convolution neural network (CNN) models to predict if a prosthesis was found to be loose in the operating room from the pre-operative XR. Each XR was pre-processed to segment the bone, implant, and bone-implant interface. A series of CNN models were built using existing, proven CNN architectures and weights optimized to our dataset. We then integrated our best performing model with historical patient data to create a final model and determine the incremental accuracy provided by additional layers of clinical information fed into the model. The models were evaluated by its accuracy, sensitivity and specificity.
Results
The CNN we built demonstrated high performance at detecting prosthetic loosening from radiographs alone. Our first model built from scratch on just the image as an input had an accuracy of 70%. Our final model which was built by fine-tuning and optimizing a publicly available model named DenseNet, combining the AP and lateral radiographs, incorporating information from the patient history, had an accuracy, sensitivity, and specificity of 98.5%, 93.9%, and 99.5% on the patients that it was trained on, and an accuracy, sensitivity, and specificity of 88.3%, 70.2%, and 95.6% on the patients it was tested on.
Conclusions
The use of machine learning (ML) can accurately detect the presence of prosthetic loosening based on plain radiographs. Its accuracy is progressively enhanced when additional clinical data is added to the loosening analysis algorithm. While this type of machine learning may not be sufficient in its present state of development as a standalone metric of loosening, it is clearly a useful augment for clinical decision making in its present state. Further study and development will be needed to determine the feasibility of applying machine learning as a more definitive test in the clinical setting.
For figures, tables, or references, please contact authors directly.
