Text

Objectives and questions: Periprosthetic Joint Infection (PJI) remains a severe complication after total hip and knee arthroplasty and its diagnosis relies on multiple clinical, laboratory, microbiological and histopathological inputs. Histopathological classification of periprosthetic tissue is largely based on manual neutrophil counting in selected high-power fields, making the diagnostic workflow labor-intensive and vulnerable to inter-observer variability. The aim of this project therefore was to develop and evaluate a deep-learning approach to distinguish septic from aseptic periprosthetic membrane samples on routine hematoxylin and eosin (H&E) whole slide images (WSI) to support standardized and rapid histopathological PJI assessment in clinical workflows.

Material and methods: In this retrospective, single-center preliminary study at a German university hospital, we collected 302 periprosthetic membrane samples from total knee arthroplasty (TKA) and total hip arthroplasty (THA) procedures performed between 2014 and 2022. 91 cases were classified as septic and 211 as aseptic according to diagnostic criteria of the European Bone and Joint Infection Society (EBJIS). H&E-stained slides (one slide per sample) were scanned at 40x magnification. Model development was performed on an NVIDIA A100 computing system using a Multiple Instance Learning (MIL) framework with a ResNet-50 backbone. To accommodate the high resolution of WSIs, multiple image patches were randomly sampled per slide, and patch-level features were aggregated via mean pooling to generate a single slide-level prediction. The model was trained using AdamW optimization and validated via stratified 5-fold cross-validation.

Results: The deep-learning model achieved an area under the receiver operating characteristic curve (AUC) of 0.940 ± 0.016. Sensitivity and specificity for detecting PJI were 0.868 ± 0.062 and 0.872 ± 0.069, respectively.

Discussion and conclusions: Deep learning applied to routine H&E whole slide images can reliably differentiate septic from aseptic periprosthetic membrane samples, reaching high discriminative performance for PJI detection. The model evaluates whole-slide morphologic patterns and is not limited to neutrophil counting in selected high-power fields, which may reduce inter-observer variability. These findings support the potential of automated WSI-based screening as a standardized, rapid tool for histopathological assessment in PJI diagnostics.