Deep learning applied to standard radiographs improves detection of implant loosening in total knee arthroplasty: A proof‐of‐concept study

TL;DR

A deep learning tool was developed to detect implant loosening in knee replacements using standard X-rays, showing better performance than traditional radiological evaluations.

Contribution

A dual InceptionV3 deep learning algorithm was proposed and shown to outperform standard radiological evaluations for implant loosening detection.

Findings

The deep learning algorithm achieved an AUC of 0.81 in detecting implant loosening.

The algorithm outperformed routine radiological evaluations (p = 0.005).

The tool showed potential to reduce analysis time and radiation exposure.

Abstract

Implant loosening is a common cause of failure in total knee arthroplasty, and for appropriate treatment, the correct diagnosis is of vital importance. Detection is challenging as radiographs only reveal obvious cases, computed tomography (CT) scans are hindered by metal artifacts, and bone scintigraphy lacks specificity, while all methods are time‐consuming and/or involve radiation exposure. This proof‐of‐concept study aimed to develop a deep learning (DL) tool to detect implant loosening using a set of two‐dimensional (2D) radiographs.

A total of 307 radiograph sets (anteroposterior and lateral) were collected, including 159 loose and 148 fixed primary knee implants (confirmed intraoperatively during revision surgery). Images were square‐cropped, centred on the implant, horizontally flipped for left knees, resized to 512 × 512 pixels and standardized for grayscale. A dual InceptionV3 DL algorithm was trained using fivefold cross‐validation. Performance was assessed using sensitivity, specificity, accuracy and receiver operating characteristic curves, with corresponding area under the curve (AUC) and compared to routine radiological reports.

The mean ± standard deviation (SD) sensitivity, specificity and accuracy over the fivefolds were 71.7% ± 10.7%, 87.0% ± 14.2% and 79.2% ± 4.0%, respectively. The overall AUC was 0.81, confidence interval (CI): [0.77–0.85]. The radiological report predictions reached a sensitivity of 51.3%, specificity of 99.2% and accuracy of 73.2%. The algorithm predicted significantly more cases correctly compared to standard radiological evaluations by a musculoskeletal specialized radiologist (p = 0.005).

This proof‐of‐concept study demonstrated the feasibility of training a deep‐learning algorithm using 2D radiographs to detect implant loosening. Notably, the algorithm outperformed routine radiological evaluations, highlighting its potential to enhance the detection of loosening. These results are promising, considering the limited dataset. This tool could be valuable for patients experiencing issues after knee replacement surgery, helping to rule out implant loosening while reducing analysis time, radiation exposure and healthcare costs.

Level III.

Figures