A machine learning model incorporating 18F-prostate-specific membrane antigen-1007 positron emission tomography/computed tomography and multiparametric magnetic resonance imaging for predicting prostate-specific antigen persistence in patients with prostate cancer after radical prostatectomy.

Abstract

Background: Although ¹⁸F-prostate-specific membrane antigen-1007 (¹⁸F-PSMA-1007) positron emission tomography/computed tomography (PET/CT) and multiparametric magnetic resonance imaging (mpMRI) are good predictors of prostate cancer (PCa) prognosis, their combined ability to predict prostate-specific antigen (PSA) persistence has not been thoroughly evaluated. In this study, we assessed whether clinical, mpMRI, and ¹⁸F-PSMA-1007 PET/CT characteristics could predict PSA persistence in patients with PCa treated with radical prostatectomy (RP).

Methods: This retrospective study involved consecutive patients diagnosed with PCa who underwent both preoperative mpMRI and PSMA PET/CT scans between April 2019 and June 2022. Scatter plots and heat maps were employed to determine the correlation of mpMRI and PSMA PET/CT features with preoperative PSA. Univariate logistic regression analyses were used assess the correlation between age, maximum Prostate Imaging-Reporting and Data System (PI-RADS) score, prostate-specific antigen density (PSAD), extracapsular extension (EPE), seminal vesicle invasion (SVI), total lesion PSMA (PSMA-TL), and PSA persistence. Multivariate logistic regression analyses were used to develop a predictive model for PSA persistence, while decision tree analysis was used to classify patients into different risk groups for easy interpretation and visualization. We divided the patient cohort into training and validation sets in an 8:2 ratio. To ensure the reliability of the model, we performed five-fold cross-validation of the validation results.

Results: Ultimately, this study included 190 patients with PCa. The median age of the patients was 69 years [interquartile range (IQR) 64-73 years]. Among the patients, 35 (18%) experienced PSA persistence following RP. Additionally, SVI was identified in 31 (16%) patients. The median values for SUVmax and PSMA-TL were 11.83 (IQR 7.44-20.89) and 41.92 (IQR 21.25-113.83), respectively. Spearman correlation analysis indicated that the preoperative PSA levels in patients with PCa were slightly correlated with the maximum standardized uptake value (SUVmax) (r=0.41; P<0.001), significantly correlated with PSMA-TL (r=0.58, P<0.001), and strongly correlated with PSAD (r=0.865, P<0.001). Multivariate logistic regression analysis showed that the independent predictors of PSA persistence were SVI on mpMRI [area under the curve (AUC)=0.63; 95% confidence interval (CI): 0.516-0.739] and PSMA-TL (AUC =0.80; 95% CI: 0.723-0.877) on PSMA PET/CT (all P values <0.05). Patients with SVI and PSMA-TL >63.38 cm³ were more likely to have PSA persistence. Decision tree analysis stratified patients into low-risk (5%), intermediate-risk (36%), and high-risk (48%) categories for PSA persistence. The model exhibited good discriminatory capability in internal validation (AUC 0.93, 95% CI: 0.850-0.930).

Conclusions: ¹⁸F-PSMA-1007 PET/CT and mpMRI parameters were proved effective in predicting PSA persistence in postoperative patients with PCa. The decision tree classification model could help clinicians to assess patients with individualized risk stratification. Patients with PSMA-TL levels below the threshold are highly likely not to have PSA persistence.