Between January 2018 and August 2020, a total of 618 men with clinical suspicion of PCa (PSA > 4 ng/ml and/or suspicious DRE results) underwent transrectal ultrasound (TRUS)-guided transperineal prostate biopsy in Ningbo First Hospital (this included both biopsy naive men as well as those with a prior negative biopsy). Inclusion criteria were PSA > 4 ng/ml and/or suspicious DRE result, and patients who underwent prostate bpMRI scan within 3 months prior to biopsy. Clinical indicators including age, tPSA, fPSA, f/tPSA, prostate volume (PV), and PSAD were also collected and measured. Exclusion criteria were treatment prior to biopsy (e.g., prostate surgery, chemotherapy, radiation therapy, or androgen deprivation therapy), bpMRI scan performed after biopsy, and incomplete data. Finally, a total of 530 patients were included in this study, 70% (n = 371) of the patients were randomly assigned to the training group, and 30% (n = 159) of the patients were assigned to the validation group. This study protocol was reviewed and approved by the Ethical Committee of Ningbo First Hospital (Ethical Approval Number: 2021B005). Written informed consent for prostate biopsy was routinely obtained before operation, and the need for informed consent of this study from patients was waived.

All bpMRI examinations were conducted with Siemens Magnetom Sonata 1.5-T superconducting MR scanner (sagittal, axial, and coronal T2WI: TR 3900 ms, TE 100 ms, FOV: 240×180, matrix: 240×160; DWI: TR 4900 ms, TE 89ms, FOV: 260×260, matrix: 100×124). The total scan time will be approximately 20 min. The routine scan protocol included T1WI, T2WI, and DWI (ADC) using body array coils (no endorectal coil). The maximum transverse diameter and anteroposterior diameter were measured in the transverse T2WI sequence, and the superoinferior diameter was measured in the sagittal T2WI sequence; PV was calculated by using the formula for prostate ellipsoid (transverse width × transverse length × longitudinal height × 0.52). Two experienced urological radiologists (Dr. ZYZ and Dr. DWR) who were blinded to pathological results and clinical information have analyzed and recorded the T2WI score and DWI score separately according to PI-RADS v.2 scoring protocol (23), and the comprehensive PI-RADS score was also performed on the bpMRI results. The different scores were discussed again until consensus was reached. Both the T2WI score and the DWI score of bpMRI results use a five-point scale based on findings on bpMRI performance. According to the PI-RADS v.2, we added the T2WI score and DWI score to obtain the Total score. The differences between PI-RADS v.2 score and Total score are summarized in Supplementary Table S1 . In order to compare the diagnostic performance of PI-RADS score and Total score of bpMRI, they were separately included in multivariate analysis of the training group to determine independent predictors for PCa and csPCa.

All biopsy operations were performed by the same experienced urological radiologist (Dr. RS) using a 7.5-MHz endocavity ultrasonic probe with color Doppler ultrasonography. Ten biopsy samples were obtained from each patient. If hypoechoic suspicious areas in peripheral zones or transition zones were found in MRI or ultrasound images, additional one to two targeted cores were also performed. All biopsy samples’ histopathological results were reported by Ningbo Pathology Center using the 2014 International Society of Urological Pathology (ISUP) modified Gleason grading system. A Gleason score of 3 + 4 or higher (ISUP ≥ 2) was defined as csPCa as the European Association of Urology Guidelines on Prostate Cancer suggested (24).

Patient characteristics were grouped according to biopsy results and summarized using descriptive statistics. Continuous variables are described as means ± standard deviations (SD) or medians and interquartile ranges (IQRs). Independent sample t-test was used for comparison between the csPCa group and the non-csPCa group. Multivariate regression analysis was used to determine the independent predictors for biopsy outcomes. Receiver operating characteristic (ROC) curve analysis and the AUC were used to examine the diagnostic performance of prediction models and clinical indicators. Spearman rank correlation was performed to analyze the relationship of categorical variables. AUC, sensitivity, specificity, negative predictive value (NPV), and positive predictive value (PPV) for each method were calculated and compared. The predictive accuracy of prediction models was validated using samples in the validation group to decrease the overfit bias. The best cutoff points for a biopsy threshold that balanced sensitivity and specificity was calculated using Youden’s J index (sensitivity + specificity − 1). ROC curves were compared by use of the method of DeLong et al. (25). Decision curve analysis (DCA) was performed and the greatest net benefit at a specific threshold probability was identified as that with the greatest clinical value. Further, Nomogram and calibration were also performed. The accuracy of the Nomogram was estimated in the training and validation groups. A p-value of <0.05 was considered statistically significant. Statistical analyses were performed with IBM SPSS (version 23.0), MedCalc (version 19.3), and R statistical software (version 4.0.4).

A total of 530 patients who met the inclusion criteria were enrolled in this study. Figure 1 is a flowchart that describes how patients and data were selected from the hospital information system (HIS). The training group and validation group consisted of 371 (70%) and 159 (30%) men, respectively, and there was no statistical difference in data characteristics between the two groups (p < 0.05). In the training group, 160 (43.1%) patients had PCa, namely, 128 (34.5%) with csPCa and 32 (8.6%) with low-grade PCa (Gleason score = 6) based on the results from both targeted and systemic biopsies ( Table 1 ). Regarding the clinical indicators, age (69.85 ± 8.07 vs. 65.86 ± 8.33 years, p < 0.001), tPSA [16.63 (8.10–24.66) vs. 10.00 (7.09–15.57) ng/ml, p < 0.001], PV [34.41 (24.70–48.01) vs. 47.42 (32.05–67.10) ml, p < 0.001], f/tPSA [0.10 (0.07–0.13) vs. 0.15 (0.11–0.18), p < 0.001], and PSAD [0.46 (0.26–0.72) vs. 0.21 (0.14–0.33) ng/ml, p < 0.001] were significantly higher in csPCa patients compared with non-csPCa (low-grade PCa and non-malignant disease) ( Table 1 ).

According to the bpMRI results, patients with PI-RADS score not less than three in the training group had a 59.5% (154/259) risk of PCa, compared to men with a lower score 5.4% (6/112). In addition, a total of 112 cases were scored as 1 and 2, namely, 2 cases of csPCa, 4 cases of low-grade PCa, and 106 cases of non-malignant lesions. A total of 101 cases were scored as 3, namely, 18 cases of csPCa, 12 cases of low-grade PCa, and 71 cases of non-malignant lesions. Regarding PI-RADS score 4 and 5, there were 108 cases of csPCa, 16 cases of non-csPCa, and 34 cases of non-malignant lesions ( Table 1 ). If the cutoff value for biopsy decision is set as PI-RADS < 4, 83.1% (177/213) of patients could avoid biopsy, at the price of missing 11.3% (20/177) csPCa cases ( Table 1 ). According to our results, it was apparent that a Total score of 6 was the optimal threshold for detecting csPCa according to ROC curve analysis and Youden’s index calculation; 87.2% (163/187) of patients with a Total score of 2–5 could avoid biopsy at the price of missing 6.9% (13/187) csPCa cases ( Table 4A ).

According to ROC curve analysis, the AUCs for diagnosing PCa and csPCa were summarized in Table 2 . Total score achieved the highest diagnostic accuracy [AUC = 0.903 (0.872–0.933), p < 0.001] in distinguishing patients with or without csPCa compared with other indicators, and had 80.5% sensitivity and 86.4% specificity. Regarding PSA derivatives, PSAD has the best diagnostic accuracy in predicting PCa [AUC = 0.734 (0.682–0.787), p < 0.001] and csPCa [AUC = 0.765 (0.711–0.819), p < 0.001] compared with other PSA derivatives ( Table 2 ). In univariable analysis, all clinical indicators were significant predictors for PCa and csPCa, except for fPSA (p > 0.05) ( Table 1 ). Spearman analysis showed that age, fPSA, tPSA, PSAD, PI-RADS score, and Total score were significantly correlated with Gleason score; Total score was most strongly associated with the malignancy of PCa ( Table 2 ).

To further construct a predictive model for biopsy decision, we performed a stepwise multivariable logistic regression analysis to eliminating the redundant variables and avoiding multicollinearity. Finally, two prediction models were established. F/tPSA, PSAD, and PI-RADS score were incorporated in model 1 (each p < 0.001); f/tPSA, PSAD, and Total score were incorporated in model 2 (each p < 0.001) for detection of PCa and csPCa, respectively ( Table 3 ). Regarding csPCa detection rate with prediction models, the AUCs of model 1 and model 2 were 0.910 (0.881–0.939, p < 0.001) and 0.931 (0.906–0.956, p < 0.001) for the training group and 0.878 (0.817–0.926, p < 0.001) and 0.910 (0.865–0.954, p < 0.001) for the validation group ( Figure 2 ). The ROC curve analysis showed that model 2 was significantly better than model 1 in both the training group and the validation group according to Delong test ( Supplementary Table S2 ). Model 2 had 88.3% sensitivity and 84.4% specificity for diagnosing csPCa. Using the model 2, 85.5% (188/220) of the study patients would have avoided biopsies and 15 csPCa (6.8%,15/220) would have been missed. Sensitivity, specificity, PPV, NPV, and AUC of PSAD; Total score; and two prediction models in the detection of csPCa are reported in Table 4 .

To assess the potential clinical benefit of the two models and clinical indicators, we performed DCA using the predicted risk in the training group and validation group. The overall utility of the decision models (i.e., PSA, PI-RADS score, Total score, model 1, and model 2) were examined ( Figure 3 and Supplemental Figure S1 ). Two models showed a higher net benefit than PSA when the threshold probability was 0.05–0.80. Model 2 showed the highest net benefit with threshold probabilities > 0.25, superior to other indicators. Assuming we choose to diagnose and treat csPCa with a predicted probability of 20%, for every 100 men using Model 2, 32 will benefit without detriment to anyone else. For every 100 men who use PSA, only 25 men benefit and without detriment to anyone else. If the threshold probability > 20%, Model 2 also has a better net benefit than other indicators such as PI-RADS score and Total score ( Figure 3 ). For example, applying a biopsy risk threshold of 40% meant that 57.1% of all men could avoid a biopsy; clinical impact curve (CIC) visually indicated the high clinical net benefit and confirmed the clinical value of model 2 ( Supplemental Figure S2 ).

In order to verify the predictive ability of Total score and Model 2, the Nomogram constructed by age, f/tPSA, PSAD, and Total score was established ( Figure 4 ). Points from each variable are added and a straight line from the total point score shows the probability of harboring csPCa. Combining the Total score with clinical factors increased the AUC for diagnosis of csPCa (AUC = 0.933). Furthermore, the calibration curves were depicted in the training group to estimate the agreement between the estimated risk of Nomogram and the actual csPCa risk. In the validation group, we verified the diagnostic performance of Nomogram in csPCa and the AUC reached 0.904. We using bootstrap method and calibration plots also confirmed the stability of the Nomogram.