A Deep Learning-Based Clinical Classification System for the Differential Diagnosis of Hip Prosthesis Failures Using Radiographs: A Multicenter Study.

Abstract

BACKGROUND Accurate and timely differential diagnosis of hip prosthesis failures remains a major clinical challenge. Radiographic examination remains the most cost-effective and common first-line imaging modality for hip prostheses, and integrating deep learning has the potential to improve its diagnostic accuracy and efficiency. METHODS A deep learning-based clinical classification system (Hip-Net) was developed to classify multiple causes of total hip arthroplasty failure, including periprosthetic joint infection (PJI), aseptic loosening, dislocation, periprosthetic fracture, and polyethylene wear. Hip-Net employed a dual-channel ensemble of 4 deep learning models trained on 2,908 routine dual-view (anteroposterior and lateral) radiographs for 1,454 patients (Asian) across 3 medical centers. An interpretive subnetwork generated spatially resolved disease probability maps. Discrimination performance and interpretability were tested in external and prospective cohorts, respectively. The correlation between model-generated individual PJI risk and inflammatory biomarkers was assessed. RESULTS Hip-Net demonstrated strong generalizability across different settings, effectively distinguishing between 5 common types of hip prosthesis failures with an accuracy of 0.904 (95% confidence interval [CI], 0.894 to 0.914) and an area under the receiver operating characteristic curve (AUC) of 0.937 (95% CI, 0.925 to 0.948) in the external cohort. The spatially resolved disease-probability maps for PJI closely aligned with intraoperative and pathological findings. The model-generated individual PJI risk scores exhibited a positive correlation with the C-reactive protein (CRP) level and erythrocyte sedimentation rate (ESR). CONCLUSIONS Hip-Net provided a clinically applicable strategy for accurately classifying and characterizing multiple etiologies of hip prosthesis failure. Such an approach is highly beneficial for providing interpretable, pathology-aligned probability maps that enhance the understanding of PJI. Its integration into clinical workflows may streamline decision-making and improve patient outcomes. LEVEL OF EVIDENCE Prognostic Level III. See Instructions for Authors for a complete description of levels of evidence.