Auxiliary diagnosis of periprosthesis joint infection by leukocyte esterase strips test using a deep learning model

The purpose of this study is to validate the efficacy of the Deep Convolutional Neural Network (DCNN) model to automatically detect Leukocyte Esterase (LE) strips in the diagnosis of Periprosthesis Joint Infection (PJI).

This study is a prospective diagnostic trial. We constructed an automatic detection system for LE strips based on the AlexNet deep convolutional model to diagnose PJI.LE strip images and white blood cell counts from 46 patients were used to train the DCNN model.96 patients were prospectively enrolled in this single-center study and underwent joint aspiratio. Based on inclusion and exclusion criteria, 18 patients were excluded; 78 patients were included for analysis, and they all agreed to participate. The ICM2018 standard was used as the gold standard to classify the included patients. One drop of synovial fluid (100 μL) from patients was applied to LE strips before and after centrifugation. The DCNN Model automatically captures the images of the strips and outputs the results. The framework of the DCNN Model includes LE strip image acquisition, preprocessing, and intelligent detection. Three observers performed LE strip colorimetric test, and the majority result was considered the final result. The clinical laboratory provided results for WBC count, PMN%, serological tests, histopathology, and microbial cultures. The diagnostic performance of the DCNN Model was analyzed using sensitivity, specificity, PPV, NPV, ROC curve, and AUC. The reliability of the DCNN Model in diagnosing PJI was assessed through a consistency test (Kappa value) with traditional PJI diagnostic methods.

The AUC of the DCNN Model was 0.92 before centrifugation and 0.91 after centrifugation. Before centrifugation, the sensitivity, specificity, PPV, and NPV of the model were 0.83, 0.97, 0.98 and 0.79, respectively. After centrifugation, they were 0.79, 1.00,1.00 and 0.76, respectively. The Kappa value for diagnosing PJI using the model before and after centrifugation was 0.82 (P<0.05). The Kappa values of the model results for PJI diagnosis compared to WBC count, PMN%, LE strip colorimetric test, ESR, CRP, and microbial culture before centrifugation were 0.82, 0.72, 0.59, 0.69, 0.64 and 0.4. After centrifugation, the Kappa values were 0.90, 0.75, 0.69, 0.67, 0.62 and 0.48.

The DCNN Model demonstrates excellent performance and high reliability in diagnosing PJI. Utilizing this model improves the objectivity of LE strip test for diagnosing PJI and mitigates the influence of external factors on the test results. Large-scale, multicenter clinical studies are necessary to further enhance the diagnostic performance of the model.