Statistical Analysis

The histopathologic results from resected lymph nodes were correlated with the results of cross-sectional imaging (MRI or CT) and ¹⁸F-rhPSMA-7 PET in a patient- and template-based manner.

The overall diagnostic accuracy of patient-level data was assessed using receiver-operating-characteristic (ROC) analyses. ROC curves were calculated for both modalities (¹⁸F-rhPSMA-7 PET and morphologic imaging). Areas under the ROC curves (AUCs) with 95% confidence intervals (CIs) were calculated and compared with each other. For patient-based analysis, the method of DeLong et al. (22) was used, and the approach proposed by Obuchowski (23) was considered for template-based analyses to account for correlations of multiple assessments within 1 patient.

To estimate sensitivities, specificities, and accuracies, a dichotomization of the semiquantitative 5-scale rating for PET and for CT or MRI was performed to conduct statistical analyses. The Youden index (sensitivity + specificity − 1) was used to determine the best cutoff for this analysis. In the patient-based analyses, exact CIs were estimated for these measures on the basis of the binomial distribution (Clopper–Pearson intervals). For the template-based analyses, logistical generalized-estimating-equation models were fitted to the data to account for the correlation of multiple observations within the same patient (24,25). For estimation of sensitivities with associated CIs, only templates with a positive histologic result were included, and the result of the diagnostic test was used as a dependent variable.

To derive estimates for the specificities, a variable indicating whether a negative test result was observed was used as a dependent variable, and only patients with a negative histopathologic result were included. Accuracy was estimated in an intercept-only model with a dependent variable that indicated whether the test result and the result of the histopathologic assessment agreed. For the generalized-estimating-equation model, an independent correlation structure was assumed. A significance level of 5% was considered for all tests. All statistical analyses were performed using the statistical software R (26), with its packages pROC (27) and geepack (28).