A deep learning framework for diagnosing periprosthetic joint infections using X-ray images: a discovery and validation study.

Researchers

Journal

Modalities

Models

Abstract

X-ray examination is the first-line imaging test for periprosthetic joint infections (PJIs). Deep learning has the potential to improve the diagnostic performance of X-ray examination for PJIs.A deep learning framework was developed for PJI diagnosis based on 1,062 X-ray images of the index prosthesis from patients who had PJI or aseptic failure. The classification network was constructed based on an ensemble of four deep learning models in a two-channel format for dual-view X-ray images. The interpret network was developed based on gradient weighted class to generate disease probability maps of individual PJI risk. The discrimination performance and disease probability maps were estimated in the validation set.This PJI deep learning technique achieved an area under the curve (AUC) of 0.913 (95% confidence interval [CI]: 0.840-0.948), sensitivity of 0.844 (95% CI: 0.768-0.861), and specificity of 0.882 (95% CI: 0.851-0.934) for PJI recognition in hip prostheses. The PJI deep learning technique achieved an AUC of 0.931 (95% CI: 0.893-0.978), sensitivity of 0.905 (95% CI: 0.806-0.942), and specificity of 0.889 (95% CI: 0.747-0.944) for PJI recognition in knee prostheses. The high-risk prosthetic regions predicted by PJI deep learning were closely tracked with intraoperative clinical and pathological findings.Deep learning provided a clinically applicable strategy for diagnosing PJI with high accuracy and robustness using routinely available X-ray images. However, the finding should be considered preliminary, the diagnosis performance might be partially attributed to prosthesis loosening, and the deep learning method is only helpful in patients already deemed suitable for revision.Copyright © 2022 Elsevier Inc. All rights reserved.