Study design and patient selection

This retrospective single-center study was conducted in accordance with the Declaration of Helsinki and was approved by the Institutional Review Board of our institution (approval No. 4965); the requirement for informed consent was waived.

We reviewed consecutive patients who underwent radical prostatectomy between January 2014 and September 2023. Initially, 153 patients were included. Twenty-five patients were excluded owing to inadequate preoperative MRI (n = 18), neoadjuvant hormone therapy (n = 3), history of transurethral resection of the prostate (n = 1), or insufficient pathological data (n = 3), resulting in 128 patients for final analysis. The patient selection flowchart is provided in Online Resource 1. Clinical data, including age and preoperative prostate-specific antigen (PSA) level, were extracted from medical records.

MRI protocol

All MRI examinations were performed using Philips MRI systems (Achieva 3 T, Ingenia 1.5T, and Ingenia Elition X 3 T; Philips Healthcare, Best, the Netherlands) with phased-array coils. All the scans met the technical recommendations of PI-RADS version 2.1. Details of the imaging process are provided in Online Resource 2.

Step 1: validation of risk stratification based on location and ADC

We first identified PI-RADS 3 lesions in the TZ on preoperative MRI and classified them based on anatomical location and ADC values. The cancer detection rate (CDR) was compared between Location-risk positive and negative lesions and between ADC-risk positive and negative lesions to evaluate risk stratification performance. This section details classification methods that served as the basis for the evaluation of the biopsy strategy described in Step 2.

Lesion identification

Prostate MRI images of the study cohort were independently reviewed according to PI-RADS v2.1 by two board-certified radiologists with 8 and 36 years of radiology experience, including 2 and 17 years of experience in genitourinary imaging (Reader 1 and Reader 2, respectively). In total, 212 TZ PI-RADS 3 lesions were identified by Reader 1 and 211 by Reader 2. After a consensus review, 218 lesions were included. Representative examples of csPCa and non-csPCa lesions are shown in Fig. 1. For each identified lesion, the maximum lesion length was measured on the ADC map by Reader 1 under the supervision of Reader 2.

Fig. 1

Representative examples of PI-RADS 3 lesions in the transition zone with corresponding MRI findings and whole-mount pathological correlation. a–e PI-RADS 3 lesion in the anterior apex in a 73-year-old man with clinically significant prostate cancer (csPCa). a T2-weighted image, b gross pathology specimen, c diffusion-weighted image (b = 2000 s/mm²), d apparent diffusion coefficient map (mean ADC, 0.960 × 10−3 mm2/s), and e schematic diagram illustrating lesion location. The black arrow indicates the lesion. f–j PI-RADS 3 lesion in the posterior base in a 60-year-old man without csPCa. f T2-weighted image, g gross pathology specimen, h diffusion-weighted image (b = 2000 s/mm²), i apparent diffusion coefficient map (mean ADC, 0.774 × 10− 3 mm2/s), and j schematic diagram illustrating lesion location. The black arrow indicates the lesion

Location-based risk stratification

The anatomical location of each lesion was determined and classified as Location-risk positive or negative according to previously published criteria [8,9,10,11,12,13,14].

A third radiologist (Reader 3, with 8 years of radiology experience, including 2 years of experience in genitourinary imaging) assessed each TZ lesion on T2-weighted images. Lesion location was recorded in the craniocaudal (base, mid-gland, apex) and anteroposterior (anterior, middle, posterior) directions, yielding nine regions. Based on prior reports describing an anterior/apical predilection of csPCa in the TZ [8,9,10,11,12,13,14], lesions located in the anterior apex, anterior mid-gland, or middle apex were predefined as Location-risk positive, whereas those in all other regions were classified as Location-risk negative (Fig. 2a).

ADC-based risk stratification

The mean ADC value of each lesion was measured and classified as ADC-risk positive or negative according to previously published criteria [20].

For each lesion, a single-slice elliptical region of interest (ROI) was placed on the ADC map at the slice showing the largest cross-sectional area of the lesion by Reader 1 under the supervision of Reader 2, with reference to corresponding axial T2-weighted images and high-b-value diffusion-weighted imaging. Both readers were blinded to histopathologic outcomes at the time of ADC measurement. The ROI was drawn to include as much of the solid lesion component as possible while avoiding the urethra, cystic or necrotic areas, hemorrhage, identifiable calcifications, and areas affected by susceptibility or motion artifacts. The mean ADC value within the ROI was recorded. Lesions with a mean ADC value < 0.800 × 10− 3 mm2/s were classified as ADC-risk positive, and those with values ≥ 0.800 × 10−3 mm2/s were classified as ADC-risk negative (Fig. 2b) [20].

Fig. 2

Risk stratification schema for PI-RADS 3 lesions in the transition zone (TZ). a Schematic illustration of the prostate showing location-based risk stratification. The TZ is partitioned into nine regions by combining craniocaudal (base, mid-gland, and apex) and anteroposterior (anterior, middle, and posterior) divisions. Regions defined as Location-risk positive—anterior apex, anterior mid-gland, and middle apex—are shaded dark gray, whereas all other regions (Location-risk negative) are shaded light gray. b Apparent diffusion coefficient (ADC)–based stratification flowchart. Lesions with mean ADC < 0.800 × 10−3 mm2/s are classified as ADC-risk positive (dark gray), whereas lesions with mean ADC ≥ 0.800 × 10−3 mm2/s are classified as ADC-risk negative (light gray). These labels are applied consistently throughout the study and used to define the biopsy selection strategies

Pathological assessment

Reader 3 reviewed clinical and pathological data and matched each lesion identified on the MRI with the corresponding lesion markings provided in the pathological report of whole-mount prostatectomy specimens. Anatomical landmarks such as the prostatic urethra, the ejaculatory ducts, and benign prostatic hyperplasia (BPH) nodules were used as references for lesion matching. After lesion matching, the pathological findings, including the Gleason score, were extracted for each lesion. CsPCa was defined as ISUP Grade Group ≥ 2 (Gleason score ≥ 3 + 4=7), and all the other lesions were classified as non-csPCa [4, 5].

Step 2: evaluation of biopsy strategies based on risk stratification

Using the risk classification established in Step 1, we designed multiple biopsy selection strategies (Fig. 3) and virtually applied them to the study cohort to evaluate their expected performance and clinical utility.

Biopsy strategy design

First, we defined a conventional strategy in which all TZ PI-RADS 3 lesions were targeted for biopsy. Although PI-RADS category 3 is not intended to mandate biopsy based solely on imaging findings, many prior studies and real-world clinical workflows have treated PI-RADS 3 lesions as biopsy targets. Accordingly, biopsy of all TZ PI-RADS 3 lesions was defined as a conventional reference strategy. In addition, we established three strategies that used anatomical location and ADC values to select target lesions (Fig. 3). The Location-based strategy targeted only Location-risk positive lesions, the ADC-based strategy targeted only ADC-risk positive lesions, and the Location + ADC-based strategy targeted lesions classified as positive by either criterion.

Fig. 3

Conceptual illustration of the four biopsy selection strategies for PI-RADS 3 lesions in the transition zone (TZ). The conventional strategy targets all the TZ PI-RADS 3 lesions. The Location-based strategy targets only Location-risk positive lesions—anterior apex, anterior mid-gland, and middle apex—shown in dark gray. The apparent diffusion coefficient (ADC)-based strategy targets only ADC-risk positive lesions (mean ADC < 0.800 × 10−3 mm2/s), shown with light-gray checkered shading. The Location + ADC-based strategy targets lesions meeting either criterion, shown with composite shading (dark gray + light-gray checkered) to represent the union of the Location- and ADC-risk positive sets. The dashed rectangle in each panel indicates the full set of TZ PI-RADS 3 lesions

Performance of four virtual biopsy strategies

The expected outcomes of each biopsy strategy were evaluated using several metrics. These included the number of lesions targeted and csPCa detected, the positive predictive value (PPV), defined as the number of csPCa lesions divided by the total number of lesions targeted for biopsy. Additionally, missed csPCa was defined as the number of csPCa lesions detected by the conventional strategy but not by a given strategy. Biopsies avoided was defined as the difference between the number of lesions targeted by the conventional strategy and those targeted by a given strategy, and its proportion was calculated. Finally, the number of biopsies avoided per missed csPCa was calculated to assess the trade-off between reducing biopsies and missing csPCa.

Decision curve analysis

Decision curve analysis was performed to compare the clinical utility of four biopsy strategies by estimating net benefit across clinically relevant threshold probabilities of 5–25% [21, 22].

Exploratory supporting analyses

Exploratory supporting analyses were performed to address methodological issues that could affect interpretation of the proposed risk stratification framework. Scanner-related ADC variability was evaluated by comparing ADC values, ADC-risk positive rates, and csPCa rates across scanner types. Potential intra-patient clustering and covariate adjustment were assessed using univariable and multivariable generalized estimating equation (GEE) logistic regression models. Reproducibility of location classification and ADC measurement was assessed in a randomly selected subset of 80 predefined lesions by comparison with the original dataset.

Statistical analysis

Statistical analyses for lesion-level comparisons were performed using JMP Pro 15 (SAS Institute Inc., Cary, NC, USA). Decision curve analysis and exploratory supporting analyses, including scanner-stratified analyses, GEE logistic regression models, and reproducibility analyses, were conducted using R (version 4.4.2; R Foundation for Statistical Computing, Vienna, Austria). The Mann–Whitney U test compared lesion length and ADC between csPCa and non-csPCa groups; Fisher’s exact test was used to assess differences in csPCa prevalence across risk categories and anatomical regions. For comparisons of CDRs across nine anatomical regions, p values from Fisher’s exact tests were adjusted using the Holm method to account for multiple comparisons. In exploratory supporting analyses, ADC values were compared across scanner types using the Kruskal–Wallis test; ADC-risk positive rates and csPCa rates were compared using Fisher’s exact test. Univariable and multivariable GEE logistic regression models with patient-level clustering, a logit link, and an exchangeable working correlation structure were used to account for potential intra-patient clustering. In the multivariable model, Location-risk classification, ADC-risk classification, age, and lesion size were included as covariates. For the reproducibility assessment, unweighted Cohen’s κ coefficients were used for categorical variables; the intraclass correlation coefficient was used for continuous mean ADC values. A p-value of < 0.05 was considered statistically significant.