Automated identification of polymorphonuclear leukocytes in hematoxylin and eosin-stained periprosthetic tissue sections using a simplified deep learning framework for periprosthetic joint infection diagnosis

Periprosthetic joint infection (PJI) is a serious complication following total hip and knee arthroplasty.Histopathological evaluation of polymorphonuclear leukocyte (PMNL) infiltration is an important diagnostic component, with commonly referenced thresholds of ≥5 and ≥10 PMNs per high-power field (HPF).However, manual PMNL counting under light microscopy is labor-intensive and subject to inter-observer variability, particularly in borderline cases.This retrospective study utilized 150 digitized HPF-level H&E-stained periprosthetic tissue section images from patients with suspected PJI at Chosun University Hospital.PMNLs were independently annotated by two trained researchers using bounding boxes, with discrepancies resolved by consensus and final verification by an attending orthopedic surgeon.A Faster R-CNN architecture with a ResNet-50 backbone and Feature Pyramid Network pretrained on ImageNet was trained for PMNL detection.On the independent test set of 30 HPF images, the model achieved an [email protected] of 0.677 for individual PMNL cell detection.Automated PMNL counts per HPF showed good agreement with manual expert counts (MAE 3.17 cells/HPF; Pearson r=0.895, p< 0.001).Threshold-based concordance was achieved in 29 of 30 HPF images at the ≥5 PMNs/ HPF threshold and 26 of 30 images at the ≥10 PMNs/HPF threshold.The proposed framework provides cell-level bounding-box detection and reproducible quantitative PMNL counting aligned with clinically relevant diagnostic thresholds.AI-assisted automated PMNL quantification may serve as an objective and scalable support tool in the diagnostic workflow of suspected PJI.