Clinically applicable optimized periprosthetic joint infection diagnosis via AI-based pathology

Abstract

Abstract Periprosthetic joint infection (PJI) is a severe complication following joint replacement surgery, often requiring complex multi-stage joint revisions or fusion, and imposing significant burdens on individuals and society as a whole. Accurate diagnosis is crucial for effective treatment. This study aimed to enhance the pathological diagnostic accuracy of PJI by standardizing an intelligent pathological diagnostic process, improving its applicability and practicality in clinical settings, and providing guidance for intelligent diagnosis of infectious diseases. We developed supervised learning models, weakly supervised learning models, and unsupervised learning PJI intelligent classification models and conducted image-level, patient-level testing, and visual verification for the first two models. The weakly supervised learning model performed nearly as well as the supervised learning model in image-level tests, achieving comparable levels of accuracy, recall rates, and ROC curves. However, in patient-level testing, the weakly supervised learning model outperformed its supervised learning counterpart. By adjusting the area threshold for the recognition regions, we significantly improved the sensitivity of PJI pathological diagnosis without compromising specificity (AUC curve area of 0.9460 for the supervised model and 0.9078 for the weakly supervised model). Based on our criteria, the existing diagnostic standard for five high-power fields in a single slide was reduced to only 3. The visualization results also revealed structural changes and loosening in the surrounding tissues, apart from localized neutrophil aggregation. Moreover, the distribution patterns of neutrophil morphology might provide clues for the diagnosis and treatment of PJI. Using an intelligent quantification and statistical approach, we successfully elevated the sensitivity of PJI pathological diagnosis to 88.42% and specificity to 92.31%. Moreover, we established unsupervised rapid auxiliary annotation models, supervised classification models, and unsupervised approximate segmentation models, thereby achieving an intelligent PJI diagnosis. Our study lays the foundation for further intelligent optimization of pathological diagnosis of other infectious diseases.