PCa was detected in 62.7% with a prevalence of csPCa in 56.3% and nsPCa in 7.4% of all cases. Detailed patient characteristics are provided in Table 1.

In univariate analysis, variables such as older age (p = 0.001), elevated PSA (p = 0.007), smaller prostate volume (p = 0,001, higher PSAD (p < 0.001), abnormal DRE and TRUS, and higher PI-RADS score (all p < 0.001) were associated with csPCa. However, on multivariable analysis, with the inclusion of all significant variables considered in the univariable analysis, only prostate volume (OR: 0.97; 95% CI: 0.95–0.99), PI-RADS score 4 (OR: 8.43; 95% CI: 4.27–16.64), and PI-RADS scores 5 (OR: 34.65; 95% CI: 13.10–91.65) were statistically significant predictors of csPCa.

The PSAD showed the best performance of all clinical parameters with an AUC of 0.70 (95%CI 0.65–0.74) for both PCa and csPCa. Statistically significant AUC differences were observed for PSAD compared to PSA (p < 0.001), DRE (p = 0.042), and age (p < 0.001). Furthermore, a significantly higher diagnostic accuracy for the PI-RADS score compared to each clinical univariate parameter for PCa (AUC 0.81 (95%CI 0.77–0.84)) and csPCa (AUC of 0.82 (95%CI 0.79–0.86)) prediction was demonstrated (all p < 0.001). ROC analyses of mpMRI-based risk models showed slightly improved diagnostic potential for PCa and csPCa detection. However, statistically significant AUC differences were detected only for the risk model by Radtke et al. for csPCa detection compared to the PI-RADS score (p = 0.019). The non-mpMRI-based ERSPC-RC3, exhibited lower diagnostic utility in detecting both PCa and csPCa (AUC 0.68 and AUC 0.76, respectively, all p < 0.01), Fig. 1 and Supplementary Fig. 1.

ROC curves for PCa and csPCa detection by multivariate risk models and mpMRI. ROC curve analysis of PI-RADS score, ERSPC-RC3, MRI-ERSPC-RC3, Radtke-RC and MSP-RC before initial prostate biopsy comparing healthy patients and men with proven PCa (A) and proven csPCa (B) is illustrated. In detail, ERSPC-RC3 (PCa: AUC 0.68, 95%CI 0.63–0.72, sensitivity 70%, specificity 61%; csPCa: AUC 0.76, 95%CI 0.72–0.80, sensitivity 65%, specificity 74%; MRI-ERSPC-RC3 (PCa: AUC 0.80, 95%CI 0.76–0.84, sensitivity 73%, specificity 73%; csPCa: AUC 0.84, 95%CI 0.81–0.87, sensitivity 80%, specificity 74%); MSP-RC (PCa: AUC 0.82, 95%CI 0.78–0.86, sensitivity 75%, specificity 78%; csPCa: AUC 0.82, 95%CI 0.79–0.86, sensitivity 78%, specificity 74%); Radtke-RC (PCa: AUC 0.82, 95%CI 0.78–0.86, sensitivity 75%, specificity 76%; csPCa: AUC 0.84, 95%CI 0.81–0.87, sensitivity 85%, specificity 65%); PI-RADS (PCa: AUC of 0.81,95%CI 0.77–0.84, sensitivity 82%, specificity 70%; csPCa: AUC 0.82, 95%CI 0.79–0.86, sensitivity 87%, specificity 68%)

To attain a more comprehensive understanding of the predictive capacities of the mpMRI within the context of modern multivariate risk models, we conducted a ROC analysis stratified based on the PI-RADS score: PI-RADS 1–2 = negative, positive = PI-RADS 3–5, and equivocal = PI-RADS 3).

The performance of the PI-RADS score and all risk models that incorporate mpMRI into their risk assessment was decreased in individuals with negative mpMRI results and those with PI-RADS 3-rated lesions. However, the diagnostic performance of all clinical variables was also diminished in men with negative mpMRI findings and those rated as PI-RADS 3. AUCs of RCs did not differ significantly, but ERSPC-RC3 and MSP-RC showed significantly greater AUC compared to PI-RADS. Of note, ERSPC-RC3 demonstrated favorable diagnostic ability for csPCa in men with negative mpMRI (AUC 0.80). In this regard, it showed superior performance relative to all other risk models within this subgroup (Fig. 2).

ROC curves for csPCa detection by multivariate risk models and mpMRI in stratified subgroups based on PI-RADS score. ROC curve analysis of PI-RADS score, ERSPC-RC3, MRI-ERSPC-RC3, Radtke-RC and MSP-RC before initial prostate biopsy comparing healthy patients and men with proven csPCa in men with negative mpMRI (negative multiparametric magnetic resonance tomography with PI-RADS (The Prostate Imaging—Reporting and Data System Version 2 (PI-RADS™ v2.1)) score 1–2, A); men with equivocal mpMRI (PI-RADS score 3, B); and men with suspicious mpMRI (PI-RADS score 3–5, C) is shown. The non mpMRI-based ERSPC-RC3 demonstrated favorable diagnostic ability for csPCa in men with negative mpMRI (AUC 0.80). It showed superior performance relative to all other risk models within this subgroup, as evidenced by a significantly higher AUC for ERSPC-RC3 in comparison to Radtke-RC (p = 0.016). In men PI-RADS 3 rated men the PSAD alone showed comparable performance to multivariate risk models (all p > 0.5: compared to Radke-RC, p = 0.538, ERSP-RC3, p = 0.850, MRI-ERSP-RC 3, p = 0.686 and MSP-RC, p = 0.934. Moreover, PSAD predicted csPCa detection better than the PI-RADS score (AUC 0.65 vs. 0.50, p = 0.020)

Despite this, the PSAD demonstrated satisfactory performance in risk stratification for csPCa in men with negative and equivocal mpMRI (AUC 0.68 (95%CI 0.56–0.80), and AUC 0.65 (95%CI 0.52–0.77), respectively). In men harboring PI-RADS 3 rated lesions the PSAD alone showed comparable performance to multivariate risk models. Moreover, PSAD predicted csPCa detection better than the PI-RADS score (AUC 0.65 vs. 0.50, p = 0.020).

In contrast, in the case of positive mpMRI findings the predictive value for csPCa by both mpMRI-based RCs (AUC 0.82–0.79) and PI-RADS score alone (AUC 0.80) showed an improved predictive ability compared to the risk assessment by clinical measures. Hereby, the AUC derived from mpMRI-based risk models and the PIRADS score displayed comparable values without statistically significant distinctions. However, they all exhibited notably superior performance when contrasted with both univariate and multivariate clinical assessments with marked statistical significance (PSAD, TRUS, DRE, PSA, age all p < 0.001; ERSPC-RC3 p < 0.002, Supplementary Table 1).

Comparative analysis revealed that the evaluation based on a quality-assured mpMRI analysis at a high-volume center (n = 419) yields superior results compared to assessments made using non-quality-assured mpMRIs (n = 146): diagnostic accuracy by RCs (AUC 0.82–0.85) and mpMRI (AUC 0.82) on-campus vs. AUC 0.74 and 0.74 in peripheral facilities for PCa; and diagnostic ability for csPCa by RCs (AUC 0.83–0.86) and mpMRI (AUC 0.83) on-campus vs. AUC 0.76–0.81 and 0.80 in peripheral facilities, respectively (Fig. 3). Descriptive statistics of both cohorts are provided in Supplementary Table 2.

ROC curves for csPCa detection by multivariate risk models and mpMRI in stratified subgroups based on mpMRI quality. ROC curve analysis of PI-RADS score, MRI-ERSPC-RC3, Radtke-RC, and MSP-RC before initial prostate biopsy comparing healthy patients and men with proven PCa (A, B) and csPCa (C, D) in men with quality-assured mpMRI (A, C) and men with mpMRI of uncertain quality (B, D) is illustrated. Absolute mean differences in AUC for mpMRI and mpMRI-derived risk models were 0.09 and 0.08 for PCa and 0.07 and 0.04 for csPCa. However, these AUC differences remained statistically insignificant for csPCa detection with an exception for Radtke-RC and MRI-ERSPC-RC3 in PCa detection (p = 0.032 and p = 0.033, respectively)

Risk stratification based solely on mpMRI demonstrated comparable results to mpMRI-based multivariate risk models, regardless image quality (Supplementary Table 3). In comparison, PCa and csPCa forecast by the non-MRI-based ERSPC-RC3 revealed a diagnostic ability of 68% (63–72%) and 76% (72–80%) in the overall cohort, respectively. Hence, using any quality mpMRI showed an improved csPCa prognostication compared to a non-MRI-based approach.

No variations were detected across the models when constidering threshold probabilities spanning from 0 to 10%. Upon reaching a 20% threshold probability, a slight decrease in the need for biopsy interventions were discernible, attributable to the employment of MRI-ERSPC-RC3, MSP-RC, or independent PI-RADS score (all < 1%). At a threshold ≥ 30% csPCa risk, only 4% of biopsies could be avoided due to the use of mpMRI-based risk models (univariate and multivariate models). A detailed description is provided in Supplementary Table 4.

In mpMRI-negative patients, a benefit in favor of using multivariate or univariate risk models is evident at lower risk thresholds of ≤ 10%, with a maximum potential reduction in biopsies to 3% when employing the ERSPC-RC3. The advantage of using risk models becomes more pronounced at higher risk probabilities. At a threshold of 10% csPCa risk, ERSPC-RC3 saves 23% of biopsies, MRI-ERSPC-RC 7%; Radtke-RC,MSP-RC and save none). At a threshold of 20% csPCa risk, consistent use of the models could have potentially avoided up to 39% of biopsies (ERSPC-RC3 39%; MRI-ERSPC-RC 28%; Radtke-RC 24%; and MSP-RC 18%, PSAD 21%).

In equivocal mpMRI findings, significant reductions in biopsies were observed when using the risk models starting at a 20% threshold probability: ERSPC-RC3 16%; MRI-ERSPC-RC 16%; Radtke-RC 0%; and MSP-RC 13%, PSAD 16%. Hence, the multivariate non-MRI-based risk calculator and the univariate risk model based on PSAD showed similar effectivity.

In patients with suspicious mpMRI results, there was an added value in utilizing risk models only at risk thresholds > 35%. The mpMRI-only approach consistently outperforms multivariate mpMRI-based risk models (representative benefit of additional csPCa detection at a risk threshold of 40%: Radtke-RC 1%; MRI-ERSPC-RC 6%; MSP-RC 6%; and mpMRI-only 7%).