External Validation of MRI-Based Prostate Cancer Risk Calculators in a Contemporary Pre-Biopsy mpMRI Cohort

1. Introduction

Prostate cancer (PCa) is the second most frequently diagnosed malignancy among men worldwide and remains a leading cause of cancer-related mortality. [1] Although prostate-specific antigen (PSA) testing has facilitated earlier cancer detection, its limited specificity results in unnecessary biopsies, overdiagnosis of clinically insignificant disease, and overtreatment, highlighting the need for more accurate pre-biopsy risk stratification. [2,3]

Over the past decade, multiparametric magnetic resonance imaging (mpMRI) has fundamentally transformed prostate cancer diagnostics. [3,4] By combining anatomical and functional imaging with standardized lesion assessment using the Prostate Imaging Reporting and Data System (PI-RADS), mpMRI improves the detection of clinically significant prostate cancer (csPCa) while reducing unnecessary biopsies. [3,4] However, MRI findings alone cannot reliably distinguish clinically significant from insignificant disease and should therefore be interpreted together with established clinical risk factors. [5]

To improve individualized risk estimation, several multivariable prediction models have been developed. [6,7,8,9] Earlier calculators, including the Prostate Cancer Prevention Trial (PCPT) and European Randomized Study of Screening for Prostate Cancer (ERSPC) models, combine routinely available clinical variables to estimate prostate cancer risk before biopsy. [6,7] Although these models performed well in their development cohorts, external validation studies demonstrated variable performance across different clinical settings. [8,9]

The integration of mpMRI into multivariable prediction models has led to MRI-based risk calculators with improved discrimination for csPCa. [10,11] Among the best validated models are the MRI-based risk calculator developed by Radtke et al. (Radtke-RC) and the MRI-integrated Rotterdam ERSPC risk calculator developed by Alberts et al. (Alberts-RC). [10,11,12,13,14] Both models combine PI-RADS assessment with established clinical variables and have demonstrated predictive performance across multiple external validation studies. [12,13,14]

A recent systematic review and meta-analysis confirmed that MRI-based risk calculators consistently improve individualized risk prediction compared with clinical models alone, although their clinical impact depends on patient selection and the diagnostic pathway in which they are applied. [15]

Both the Radtke-RC and Alberts-RC have demonstrated good predictive performance in previous external validation studies. [12,13,14] However, direct comparison of their discriminative performance and clinical utility within the same cohort of patients selected for biopsy after pre-biopsy mpMRI may provide further insight into their relative performance in clinical practice.

The aim of the present study was to externally validate the Radtke-RC and Alberts-RC in a contemporary single-centre cohort of patients undergoing pre-biopsy mpMRI and subsequent prostate biopsy.

2. Materials and Methods

This retrospective single-centre study was conducted at the Department of Urology, Paracelsus Medical University Salzburg, Austria. The study was approved by the Ethics Committee of the Federal State of Salzburg (approval number: 1081/2025) and was conducted in accordance with the Declaration of Helsinki. Consecutive patients who underwent pre-biopsy multiparametric magnetic resonance imaging (mpMRI) followed by prostate biopsy between October 2023 and September 2025 were retrospectively identified. Clinical, radiological, and pathological data were extracted from the institutional electronic medical records and pseudonymized before analysis.

Collected variables included age, pre-biopsy serum prostate-specific antigen (PSA), prostate volume, digital rectal examination (DRE) findings, previous negative biopsy status, and PI-RADS score. Histopathological data included biopsy outcome and Gleason score/ISUP Grade Group.

Patients were eligible if clinical, radiological, and pathological data required for risk calculation and outcome assessment were available. Patients with missing predictor variables or pathological outcomes were excluded. Complete-case cohorts were defined separately for the Radtke-RC and Alberts-RC according to the predictor variables required by each model.

All mpMRI examinations were interpreted by experienced radiologists, and suspicious lesions were classified according to PI-RADS version 2.1.4 For patients with multiple lesions, the highest PI-RADS score was used for subsequent analyses. All patients underwent combined MRI-targeted and systematic fusion-guided prostate biopsy via a transperineal or transrectal approach according to institutional practice. Histopathological evaluation was performed according to the International Society of Urological Pathology (ISUP) grading system of 2019. [16]

The primary endpoint was clinically significant prostate cancer (csPCa), defined as ISUP Grade Group ≥2 on prostate biopsy, consistent with the outcome definition used in the original model studies.10,11 ISUP Grade Group 1 tumors were classified as clinically insignificant prostate cancer. Biopsy histopathology served as the reference standard for outcome classification, irrespective of subsequent radical prostatectomy findings.

Predicted probabilities of csPCa were calculated using the MRI-integrated risk calculators developed by Radtke et al. [10] and Alberts et al. [11] The Radtke-RC was implemented according to the published logistic regression equation.10 For the Alberts-RC, predicted probabilities were obtained using the publicly available online Rotterdam ERSPC risk calculator provided by the Prostate Cancer Research Foundation (Stichting Wetenschappelijk Onderzoek Prostaatkanker [SWOP]). [11] The Alberts-RC was therefore evaluated as a black-box prediction model because the underlying regression coefficients were not publicly available.

Continuous variables are presented as mean ± standard deviation (SD), and categorical variables as frequencies and percentages. Discrimination was assessed using receiver operating characteristic (ROC) analysis, and the area under the ROC curve (AUC) with corresponding 95% confidence intervals (CI) was calculated. The discriminative performance of the Radtke-RC and Alberts-RC was directly compared using DeLong’s test for paired ROC curves in the complete-case cohort (n = 114). [17]

Clinical utility was assessed using decision curve analysis (DCA) across the evaluated range of threshold probabilities by comparing the net benefit of each risk calculator with the default strategies of biopsying all patients and performing no biopsy. [18] Net-reduction curves were additionally used to quantify the potential reduction in unnecessary biopsies relative to a biopsy-all strategy. For descriptive interpretation, threshold probabilities of 12%, 20%, and 30% were evaluated, with the 12% threshold additionally included to facilitate comparison with the previous external validation by Pallauf et al. [19] The corresponding number needed to assess (NNA) was calculated to express the number of patients who would need to be assessed using the risk calculator to avoid one unnecessary biopsy. For complementary interpretation, incremental net benefit relative to the biopsy-all strategy was also expressed as net true-positive equivalents per 100 patients. The reciprocal of the incremental net benefit was used to derive the number needed to assess to obtain one additional net true-positive equivalent; these values represent decision-analytic equivalents rather than additionally detected csPCa cases. The selected thresholds were used for descriptive interpretation and were not considered prespecified clinical biopsy cut-offs.

Statistical analyses were performed using R version 4.4.0, and statistical significance was defined as a two-sided p value <0.05. Given that no adjustment for multiple testing was performed, inferential results should be interpreted as exploratory and hypothesis-generating.

3. Results

A total of 171 patients were screened for eligibility. Of these, 164 patients had sufficient data for external validation of the Radtke-RC and constituted the Radtke-RC validation cohort. Within this cohort, 114 patients additionally had complete predictor data required for validation of the Alberts-RC and therefore constituted the Alberts-RC validation cohort. Thus, 7 of 171 patients were not evaluable for the Radtke-RC. For the Alberts-RC, an additional 50 patients lacked the required predictor data, resulting in a total of 57 patients who were not evaluable and 114 patients who were included in the Alberts-RC validation. The 114 patients with complete data for both models also constituted the cohort used for direct head-to-head comparison. Baseline characteristics according to model availability are presented in Table 1.

Figure 1. Flowchart of patient selection and formation of the model-specific analysis cohorts: Of 171 patients screened for eligibility, 164 had sufficient data for validation of the Radtke-RC. Among these, 114 patients had complete predictor data required for validation of the Alberts-RC and constituted the complete-case cohort used for direct head-to-head comparison of both models. Abbreviations: RC, risk calculator.

Figure 1. Flowchart of patient selection and formation of the model-specific analysis cohorts: Of 171 patients screened for eligibility, 164 had sufficient data for validation of the Radtke-RC. Among these, 114 patients had complete predictor data required for validation of the Alberts-RC and constituted the complete-case cohort used for direct head-to-head comparison of both models. Abbreviations: RC, risk calculator.

Clinically significant prostate cancer (csPCa) was detected in 62 of 164 patients (37.8%) evaluable for the Radtke-RC and in 41 of 114 patients (36.0%) evaluable for the Alberts-RC. Among patients evaluable for the Radtke-RC, csPCa detection rates were 23.5% (4/17) for PI-RADS 3, 29.0% (27/93) for PI-RADS 4, and 59.6% (31/52) for PI-RADS 5. Corresponding rates among patients evaluable for the Alberts-RC were 28.6% (4/14), 25.4% (17/67), and 62.5% (20/32), respectively.

Among the 62 patients diagnosed with csPCa, biopsy-specific detection could be evaluated in 57 (91.9%). Of these, csPCa was detected by both MRI-targeted and systematic biopsy in 37 patients (64.9%), exclusively by MRI-targeted biopsy in 10 patients (17.5%), and exclusively by systematic biopsy in 10 patients (17.5%). In the remaining five patients, biopsy-specific detection could not be reliably determined.

In the full Radtke-RC validation cohort (n = 164), the Radtke-RC achieved an AUC of 0.777 (95% CI 0.702–0.852). In the complete-case cohort used for direct head-to-head comparison (n = 114), the Radtke-RC achieved an AUC of 0.785 (95% CI 0.696–0.874), compared with 0.767 (95% CI 0.672–0.861) for the Alberts-RC. Direct comparison using DeLong’s test demonstrated no statistically significant difference in discriminative performance between the two models (p = 0.318).

Figure 2. Receiver operating characteristic curves for the head-to-head comparison of the Radtke-RC and Alberts-RC for prediction of clinically significant prostate cancer: Both curves are based on the complete-case cohort (n = 114). The Radtke-RC achieved an AUC of 0.785 (95% CI 0.696–0.874) and the Alberts-RC an AUC of 0.767 (95% CI 0.672–0.861). The models were compared using the paired DeLong test (p = 0.318). The diagonal reference line represents no discriminative ability (AUC = 0.50). Abbreviations: ROC, receiver operating characteristic; AUC, area under the curve; CI, confidence interval; csPCa, clinically significant prostate cancer; RC, risk calculator.

Figure 2. Receiver operating characteristic curves for the head-to-head comparison of the Radtke-RC and Alberts-RC for prediction of clinically significant prostate cancer: Both curves are based on the complete-case cohort (n = 114). The Radtke-RC achieved an AUC of 0.785 (95% CI 0.696–0.874) and the Alberts-RC an AUC of 0.767 (95% CI 0.672–0.861). The models were compared using the paired DeLong test (p = 0.318). The diagonal reference line represents no discriminative ability (AUC = 0.50). Abbreviations: ROC, receiver operating characteristic; AUC, area under the curve; CI, confidence interval; csPCa, clinically significant prostate cancer; RC, risk calculator.

Decision curve analysis demonstrated threshold-dependent clinical utility for both MRI-integrated risk calculators. The Radtke-RC provided greater net benefit than the biopsy-all strategy from a threshold probability of approximately 9% and remained superior up to approximately 69%, whereas the Alberts-RC exceeded biopsy-all from approximately 18% to 56%.

At a risk threshold of 12%, corresponding to weighting the consequences of missing csPCa approximately 8 times more strongly than those of performing an unnecessary biopsy, the net benefit of the Radtke-RC was 0.274 compared with 0.272 for biopsy-all. This absolute net-benefit gain of 0.0012 was equivalent to approximately 0.12 additional net true-positive outcomes per 100 patients, or approximately 836 patients needing to be assessed to obtain one additional net true-positive equivalent. The corresponding net reduction was 0.9 unnecessary biopsies per 100 patients, equivalent to approximately 114 patients needing to be assessed to avoid one unnecessary biopsy. At the same threshold, the Alberts-RC had a lower net benefit than biopsy-all (0.256 vs. 0.272) and therefore provided no clinical benefit over the biopsy-all strategy.

At a threshold probability of 20%, where the consequences of missing csPCa are weighted four times more strongly than those of an unnecessary biopsy, both calculators provided greater net benefit than biopsy-all. Net benefit was 0.219 for the Radtke-RC and 0.211 for the Alberts-RC, compared with 0.200 for biopsy-all. These differences corresponded to approximately 2.0 and 1.1 additional net true-positive equivalents per 100 patients, respectively, or approximately 51 and 91 patients needing to be assessed to obtain one additional net true-positive equivalent. The corresponding net reductions were 7.9 and 4.4 unnecessary biopsies per 100 patients, translating into approximately 13 and 23 patients, respectively, needing to be assessed to avoid one unnecessary biopsy.

At a threshold probability of 30%, corresponding to weighting the consequences of missing csPCa approximately 2 times more strongly than those of an unnecessary biopsy, the net benefit was 0.140 for the Radtke-RC and 0.132 for the Alberts-RC, compared with 0.085 for biopsy-all. This represented approximately 5.5 and 4.6 additional net true-positive equivalents per 100 patients, respectively, corresponding to approximately 18 and 22 patients needing to be assessed to obtain one additional net true-positive equivalent. The respective net reductions were 12.9 and 10.8 unnecessary biopsies per 100 patients, equivalent to approximately 8 and 9 patients needing to be assessed to avoid one unnecessary biopsy.

Overall, clinical utility was strongly dependent on the selected threshold probability. At the lower threshold of 12%, additional benefit over biopsy-all was minimal for the Radtke-RC and absent for the Alberts-RC. At thresholds of 20% and 30%, both calculators provided greater net benefit than biopsy-all, with clinically more tangible gains in terms of avoiding unnecessary biopsies. The Radtke-RC showed numerically greater net benefit than the Alberts-RC at both representative thresholds.

Figure 3. Decision curve analysis of net benefit: Curves are based on the complete-case cohort (n = 114) and compare the Radtke-RC and Alberts-RC with the default strategies of biopsying all patients and performing no biopsy. Threshold probabilities are displayed across the full 0–100% axis; decision-curve quantities were calculated from 1% to 99%, as the weighting term pₜ/(1−pₜ) is undefined at 100%. Abbreviations: RC, risk calculator; csPCa, clinically significant prostate cancer.

Figure 3. Decision curve analysis of net benefit: Curves are based on the complete-case cohort (n = 114) and compare the Radtke-RC and Alberts-RC with the default strategies of biopsying all patients and performing no biopsy. Threshold probabilities are displayed across the full 0–100% axis; decision-curve quantities were calculated from 1% to 99%, as the weighting term pₜ/(1−pₜ) is undefined at 100%. Abbreviations: RC, risk calculator; csPCa, clinically significant prostate cancer.

Figure 4. Decision curve analysis of net reduction in unnecessary biopsies: Net reduction is shown relative to the biopsy-all strategy and is expressed as the number of unnecessary biopsies avoided per 100 patients based on the corresponding difference in net benefit. The biopsy-all strategy has a net reduction of zero by definition; the biopsy-none strategy is additionally shown. Both model curves are based on the complete-case cohort (n = 114). Threshold probabilities are displayed across the full 0–100% axis; calculations were performed from 1% to 99%, as the decision-curve weighting term is undefined at 100%. Abbreviations: RC, risk calculator; csPCa, clinically significant prostate cancer.

Figure 4. Decision curve analysis of net reduction in unnecessary biopsies: Net reduction is shown relative to the biopsy-all strategy and is expressed as the number of unnecessary biopsies avoided per 100 patients based on the corresponding difference in net benefit. The biopsy-all strategy has a net reduction of zero by definition; the biopsy-none strategy is additionally shown. Both model curves are based on the complete-case cohort (n = 114). Threshold probabilities are displayed across the full 0–100% axis; calculations were performed from 1% to 99%, as the decision-curve weighting term is undefined at 100%. Abbreviations: RC, risk calculator; csPCa, clinically significant prostate cancer.

4. Discussion

The present study externally validated two established MRI-integrated prostate cancer risk calculators in a contemporary cohort of patients undergoing pre-biopsy mpMRI and yielded three principal findings. First, both the Radtke and Alberts Prostate Cancer Risk Calculators demonstrated acceptable discriminative performance for predicting clinically significant prostate cancer (csPCa), with no statistically significant difference between the two models. According to the AUC classification applied by Blas et al., AUC values between 0.7 and <0.8 indicate acceptable discrimination; accordingly, the observed AUCs of 0.785 for the Radtke-RC and 0.767 for the Alberts-RC in the head-to-head cohort fall within this range. [20] Second, decision curve analysis demonstrated threshold-dependent clinical utility for both calculators, with both models providing greater net benefit than a biopsy-all strategy across substantial parts of the evaluated threshold range. Third, neither calculator consistently outperformed the other in terms of clinical utility, emphasizing the importance of the selected decision threshold when applying these models in clinical practice. [18]

The discriminative performance observed in the present study is consistent with the original development studies of both MRI-integrated risk calculators. [10,11] Radtke et al. demonstrated that combining PI-RADS assessment with established clinical variables improved prediction of csPCa compared with clinical parameters alone, whereas Alberts et al. reported similar improvements following integration of mpMRI into the Rotterdam ERSPC risk model. [10,11] Subsequent external validation studies evaluated these prediction models across different patient populations and biopsy strategies. [12,13,14] Our findings therefore further support the generalizability of both MRI-integrated risk calculators in contemporary clinical practice. [10,11,12,13,14,15]

An important characteristic of the present study is its selected pre-biopsy mpMRI population. All patients underwent pre-biopsy mpMRI followed by combined MRI-targeted and systematic fusion-guided prostate biopsy, reflecting an MRI-based diagnostic pathway in which biopsy decisions had already been influenced by MRI findings before application of the risk calculators. This preselection may have affected the additional clinical utility of MRI-integrated prediction models and should be considered when interpreting the threshold-dependent net benefit observed in the present study. [5,15]

Our findings should therefore be interpreted as complementary rather than contradictory to previous validation studies. [12,13,14,15] Of particular relevance, Pallauf et al. externally validated the same two MRI-integrated risk calculators in a cohort consisting exclusively of men with a previous negative systematic biopsy and similarly demonstrated that the clinical value of the calculators depended strongly on the selected risk threshold. [19] However, their cohort differed substantially from the present population with respect to previous biopsy status, PI-RADS distribution, and csPCa prevalence. [19] Whereas all patients in the Pallauf cohort had undergone a previous negative biopsy, our cohort predominantly comprised patients with suspicious MRI findings undergoing MRI-targeted biopsy. Such differences in case mix and disease prevalence are particularly relevant for external validation and may explain differences in both discrimination and decision-analytic performance across studies. [15,19]

These findings are supported by previous comparative studies demonstrating that several MRI-integrated risk calculators achieve similar discriminative performance when evaluated in contemporary MRI-based diagnostic pathways. [5,21] Furthermore, the recently published systematic review and meta-analysis by O’Toole et al. confirmed that MRI-based risk calculators consistently improve individualized prediction of csPCa compared with clinical models alone, although their clinical impact depends on patient selection and the diagnostic pathway in which they are applied. [15] Our findings support this evidence by demonstrating that both calculators retained acceptable discriminative performance while their clinical utility varied according to the selected decision threshold.

From a clinical perspective, the present decision curve analysis illustrates why discrimination alone is insufficient to determine the usefulness of a prediction model. [18] The threshold probability incorporates the relative consequences assigned to missing csPCa and performing an unnecessary biopsy. Thus, at a threshold of 20%, missing one csPCa is weighted four times more strongly than performing an unnecessary biopsy, whereas at 30% this relative weighting decreases to approximately 2.3:1. At the lower threshold of 12%, where missing csPCa is weighted approximately 7.3 times more strongly, the incremental net benefit of the Radtke-RC over biopsy-all was marginal, while the Alberts-RC provided no incremental benefit. In contrast, at 20%, both models provided positive incremental net benefit and reduced unnecessary biopsies by 7.9 and 4.4 per 100 patients for the Radtke-RC and Alberts-RC, respectively. At 30%, these reductions increased to 12.9 and 10.8 per 100 patients. Accordingly, approximately 13 and 23 patients would need to be assessed at the 20% threshold, and approximately 8 and 9 patients at the 30% threshold, to avoid one unnecessary biopsy using the Radtke-RC and Alberts-RC, respectively.

These results highlight the central trade-off underlying application of the calculators. At low threshold probabilities, where avoiding missed csPCa is prioritized strongly, a biopsy-all strategy is difficult to improve upon because nearly all patients are considered candidates for biopsy. As the threshold increases and greater relative importance is assigned to avoiding unnecessary biopsy, the risk calculators provide progressively more opportunity for risk stratification. This pattern is conceptually consistent with the findings of Pallauf et al., who likewise observed limited additional true-positive benefit at low thresholds but substantially greater potential for identifying patients in whom biopsy might be avoided. [19] Importantly, however, higher thresholds necessarily imply a greater willingness to accept missed csPCa. The increasing reduction in unnecessary biopsies should therefore not be interpreted as evidence that higher thresholds are inherently preferable, but rather as illustrating the trade-off between cancer detection and biopsy avoidance captured by decision curve analysis. [18,19]

Taken together, these findings suggest that the principal clinical contribution of MRI-integrated risk calculators in an already MRI-selected biopsy population may not be a substantial increase in csPCa detection, but rather improved identification of patients in whom biopsy could reasonably be reconsidered. The choice of threshold may consequently be at least as important as the choice between the two calculators themselves. Because neither model consistently dominated the other across the evaluated threshold range, our findings do not support a universally preferable calculator or a single optimal biopsy threshold. Instead, threshold selection should reflect the clinical context and the relative importance assigned by patients and clinicians to avoiding unnecessary biopsy versus minimizing the risk of missed csPCa. [18] MRI-integrated risk calculators should therefore be regarded as complementary decision-support tools for individualized and shared biopsy decision-making rather than as stand-alone determinants of biopsy indication. [15,18]

Several limitations should be acknowledged. First, this was a retrospective single-centre study, which may limit the generalizability of the findings. Second, the relatively small sample size, particularly within the complete-case cohort used for direct comparison of both models, may have limited statistical power to detect small differences between the prediction models. Third, no centralized re-review of MRI examinations was performed. Finally, the study included a highly selected population of patients who underwent pre-biopsy mpMRI followed by prostate biopsy. This selection limits the generalizability of our findings to unselected screening or pre-MRI populations and may have influenced both model performance and estimates of clinical utility. [5,15]

5. Conclusions

Both the Radtke and Alberts MRI-integrated risk calculators demonstrated acceptable and comparable discrimination for csPCa, with neither model showing a clear overall advantage. Clinical utility was strongly threshold-dependent, with limited incremental benefit over biopsy-all at low thresholds but increasing potential to reduce unnecessary biopsies at higher thresholds. These findings suggest that the principal value of MRI-integrated risk calculators in an already MRI-selected population lies in supporting individualized biopsy decisions. Threshold selection should therefore reflect the clinical trade-off between avoiding unnecessary biopsy and minimizing missed csPCa. Further prospective multicentre validation is warranted to define clinically appropriate thresholds.