Introduction:

The implementation of multiparametric magnetic resonance imaging (mpMRI) prior to prostate biopsy has improved detection of clinically significant prostate cancer through use of mpMRI–transrectal ultrasound (TRUS) fusion-guided biopsies. As use of mpMRI increases, providers and patients have access to imaging findings which are not considered in traditional prediction tools including the recently developed Prostate Biopsy Collaborative Group (PBCG) risk calculator. Therefore, we aimed to develop a prostate cancer risk calculator incorporating mpMRI findings, compare its ability to predict prostate cancer detection with the PBCG risk calculator, and estimate trade-offs in the ability to avoid unnecessary biopsies.

Methods:

A prospective cohort of patients received mpMRI prior to prostate biopsy (January 2015-June 2020) as part of the Prospective Loyola University mpMRI (PLUM) Prostate Biopsy Cohort. Men receiving their first mpMRI without a previous diagnosis of prostate cancer were included. Clinical variables included age, race, family history of prostate cancer, digital rectal examination, prior negative biopsy, PSA, prostate volume, and presence of mpMRI lesion (scored by Prostate Imaging-Reporting and Data System (PI-RADS) version 2.0). Patients were separated into training (two-thirds) and validation (one-third) cohorts. First, multivariable logistic regression models evaluated standard clinical variables followed by addition of PI-RADS and PSA density (PSAD) for detection of prostate cancer. Second, a multinomial logistic regression model was constructed with outcomes of no cancer, clinically insignificant cancer (Gleason score 3+3), and clinically significant cancer (Gleason score ≥3+4). Receiver operating characteristics (ROC) and calibrations curves were evaluated and compared to the PBCG model.

Results:

A total of 900 patients were separated into training (N=600) and validation (N=300) cohorts. In the training cohort, 254 (42.3%) of patients were diagnosed with prostate cancer and 178 (29.7%) represented clinically significant disease. Addition of PI-RADS improved model performance (AUC: 0.826 vs. 0.729,p<0.001). Replacing log-transformed PSA with log-transformed PSAD calculated from mpMRI prostate volumes additionally improved discrimination (AUC: 0.859 vs. 0.826,p<0.001) and outperformed the PBCG risk calculator (AUC: 0.859 vs. 0.668,p<0.001) for detection of prostate cancer. The final PLUM risk calculator from multinomial logistic regression outperformed PBCG for detection of any cancer (AUC: 0.854 vs. 0.668,p<0.001) and clinically significant cancer (AUC: 0.873 vs. 0.665,p<0.001) with similar findings in the validation cohort. The PBCG model was prone to overprediction while the PLUM model was well calibrated. At a cost level of missing 7.5% of clinically significant cancers, the PLUM model would avoid 39.6% of biopsies compared to 15.7% for PBCG.

Conclusion:

The addition of mpMRI to standard clinical variables significantly improved the ability to accurately predict prostate cancer detection, including clinically significant prostate cancer, compared to the PBCG risk calculator. Discrimination was augmented by incorporating PI-RADS score and PSAD based on mpMRI prostate volume which are readily accessible data points as more patients undergo mpMRI. A large proportion of biopsies could be avoided using the PLUM risk calculator in shared-decision making to inform the decision for prostate biopsy.

Funding: Siemens