The Institutional Review Board of Center 1 approved the study and granted a waiver for acquiring patient informed consent. In this institution, ADC measurements described below were performed de novo for the purpose of the study. In Center 2, a fully anonymized database of ADC values of prostate lesions was already collected for the purpose of different research, with use for any future retrospective study granted by local Ethical Review Board approval and previously acquired written informed consent. The period of patient inclusion was March 2020–September 2022 in Center 1 and May 2021–April 2023 in Center 2.

In both institutions, we included ≥ 18-year-old men who consecutively underwent MRI followed by prostate biopsy because of at least one PI-RADS ≥ 3 lesion or higher clinical risk despite a negative examination (PI-RADS ≤ 2). Indications of MRI were increased prostatic specific antigen (PSA) serum level (≥ 3.0 ng/mL in two serial samples) and/or suspicious digital rectal examination (DRE). All the included men were of Caucasian ethnicity. Criteria for auctioning prostate biopsy in men with negative MRI were not standardized but rather at the urologists’ discretion based on PSA and PSA density values, PSA kinetics, DRE, family history, and prior negative biopsy, if any. Exclusion criteria are illustrated in the study flowchart (Fig. 1). Ongoing therapy with 5-alfa reductase inhibitors was among the exclusion criteria to prevent reduced conspicuity on high b-value images [19].

Study flowchart. No men were excluded because of the absence of measurable findings in Center 2. BPH benign prostatic hyperplasia, bpMRI biparametric magnetic resonance imaging, mpMRI multiparametric magnetic resonance imaging, PZ peripheral zone, TURP Transurethral resection of the prostate

All the included men received 12-core systematic prostate biopsy plus target biopsy on all the suspicious MRI findings (PI-RADS ≥ 3) prompted in the original MRI report. Lesions reported as PI-RADS 1 or 2 underwent target biopsy in no cases. Additional details on the biopsy procedure are shown in the Supplementary Material. Histological analysis of the biopsy samples by referring uropathologists was the standard of reference for csPCa, which was assumed to be an ISUP grading group ≥ 2 cancer [20].

Target biopsy included four cores (two in-target and two perilesional cores) in Center 1 and Center 2. In both centers, the procedure was performed under local anesthesia by one of a pool of experienced urologists, using fusion ultrasound-mpMRI guidance (Applio 300 platform, Toshiba/Canon in Center 1; Uronav system by Philips Healthcare in Center 2). The biopsy route was transperineal in Center 1 and transrectal in Center 2.

MRI examinations were performed on one of several 3.0-T magnets (Achieva, Philips Medical Systems, in Center 1 and several 3-T units, mostly Prisma fit, Siemens Healthineers, in Center 2) using a surface coil.

Both centers utilized a dual DWI sequence approach, using the vendor’s software to perform a linear regression of signal intensity versus the b-values of the first DWI sequence with a maximum b of at least 1000 s/mm2 to generate the ADC map. Technical acquisition details are shown in Supplementary Tables 1 and 2.

In Center 1, one reader qualified as an expert according to reference criteria [21] replicated the image analysis strategy previously adopted in Center 2 by one different radiologist with comparable experience. Center 1 readings were performed retrospectively compared to MRI and prostate biopsy. Readers were blinded to clinical history and biopsy results but not to the MRI report and were asked to measure the ADC of any previously reported PI-RADS 1–5 lesion in the PZ using a dedicated console (SuitEstensa, Esaote, in Center 1 and Impax, Agfa HealthCare, in Center 2). In both centers, previously reported lesions were identified for ADC measurement using the PI-RADS sectorial map and key images appended to the description made in the original report. Reflecting the EUSOBI principles to obtain the DLs [15], the ADC was measured by placing the largest possible region of interest (ROI) on the most hypointense part of a lesion in the ADC map, having care this corresponded to a visible signal alteration on the high b-value image and/or on dynamic contrast-enhanced imaging, if any. In the case of uncertainty about the most hypointense part of the lesion, multiple ROIs were placed, and the one with the lowest ADC value was selected for analysis. Notably, MRI examinations classified as PI-RADS 1 showing no focal lesions at all in the PZ were excluded from analysis (Fig. 1), based on the assumption that DLs have limited significance in the absence of measurable findings. Thus, ADC was measured on “PI-RADS 1” measurable lesions, reported as such because of their clearly benign appearance (e.g., ectopic nodules of the transition zone). This led to build DLs over the whole spectrum of measurable lesions.

DLs were derived from the ADC values by modifying the step-by-step methodology used by Bickel et al [18]. First, we plotted all the collected ADC values against the cumulative frequency of csPCa, testing the correlation with Kendall’s tau index. Second, we run a receiving operating characteristics (ROC) analysis to identify the ascending ADC value ranges corresponding to the following intervals of sensitivity and specificity in assessing csPCa, which were, in turn, assumed to represent as many DLs: sensitivity < 60% and specificity > 90% (very low diffusion level (VL-DL)); sensitivity 60–75% and specificity 85–90% (low diffusion level (L-DL)); sensitivity 75–95% and specificity 70–85% (intermediate diffusion level (I-DL)); sensitivity > 95% and specificity ≤ 70% (high diffusion level (H-DL)). We then calculated the actual prevalence of benign lesions, ISUP ≥ 1 cancers, and ISUP ≥ 2 cancers on a per-DLs basis and used Kendall’s tau index to assess the correlation between DLs and the observed ISUP grading groups.

The ADC threshold found at ROC analysis according to the Youden index was used in further analysis as the “absolute ADC cutoff” mentioned below.

As the Shapiro-Wilk test showed a non-normal distribution of continuous variables, they were summarized by reporting the median values with the interquartile range (IQR). However, we also showed mean ± standard deviation values to facilitate the comparison with previous literature. The Kruskal-Wallis test was used to compare the ADC values across benign lesions, ISUP 1 and ISUP ≥ 2 cancers, using the u Mann–Whitney test for pairwise comparisons. The alfa level was 0.05 with the Bonferroni correction when applicable.

To assess whether DLs improve the diagnostic performance of MRI-informed biopsy decisions, we compared six different biopsy strategies against strategy 1, i.e., sampling all PI-RADS ≥ 3 lesions (Fig. 2). Biopsy strategies 2–4 were supposed to adjust the PI-RADS categorization with the absolute ADC cutoff found at ROC analysis. In contrast, strategies 5–7 were supposed to adjust the PI-RADS categorization with DLs. We calculated the sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) for csPCa at a lesion-level, assuming a fixed comparable threshold for auctioning prostate biopsy, i.e., PI-RADS 3 category (strategy 1) or PI-RADS 3-adjusted category (strategies 2–7). Biopsy targets above the threshold were categorized as “true positive” if csPCa was found in at least one core of the target biopsy or the systematic biopsy performed in an adjacent quadrant. Biopsy targets below the threshold were defined as “true negative” if there was no csPCa in any target biopsy core or systematic biopsy core from the adjacent quadrants. Given the assumption of a fixed threshold for biopsy, we did not run ROC analysis to compare the strategies.

Biopsy decisions based on PI-RADS version 2.1 alone (strategy 1), PI-RADS version 2.1 adjusted with an absolute apparent diffusion coefficient (ADC) cutoff of 1.0 × 10−3 mm2/s (strategy 2 to 4), and PI-RADS version 2.1 adjusted with the Diffusion Levels (DLs)

Finally, we used decision curve analysis to establish which biopsy strategy showed the greatest clinical utility in terms of net benefit, i.e., the advantage of detecting true positives adjusted for the harm of false positives against a “treat none” strategy (biopsying no lesions) and a “treat all” strategy (biopsying all lesions).