A Feasibility study of a two-step ADC-PSAD-GWR rule for stratifying prostate lesions prior to biopsy

Abstract

Objectives

To test the feasibility of a two-step rule integrating apparent diffusion coefficient (ADC), PSA density (PSAD), and gadolinium wash-in/wash-out rate (GWR) in stratifying prostate lesions for biopsy decision-making, and to compare its performance with PI-RADS v2.1 within a retrospective single-center cohort.

Methods

This retrospective case series included 110 patients (mean age 70 ± 9 years) who underwent 3.0-T mpMRI and prostate surgery. ADC, PSAD, and GWR were measured from index lesions. Optimal cutoffs (ADC 0.747 × 10⁻³ mm²/s, PSAD 0.344 ng/mL², GWR 9.454%) were derived by 5-fold cross-validation. A two-step rule was constructed: step 1 assigned lesions to four ADC-PSAD quadrants (Q1–Q4); step 2 refined risk using GWR. Performance was compared descriptively with PI-RADS.

Results

The four-quadrant framework identified Q2 (low ADC/high PSAD) as the malignant-dominant quadrant and Q3 (high ADC/low PSAD) as predominantly benign. GWR refinement defined a subgroup (Q3 with low GWR) that contained no clinically significant prostate cancer (csPCa) among the 20 patients assigned. In the sensitivity analysis (all-grade PCa vs. benign), the same subgroup contained zero cancers. Compared with PI-RADS, the two-step rule avoided 35.1% (13/37) of unnecessary biopsies in PI-RADS ≥ 4 and detected all occult csPCa in PI-RADS ≤ 3 (12/12). Overall csPCa detection was 97.6% (41/42). Sensitivity analysis yielded consistent findings: unnecessary biopsy avoidance of 46.4% (13/28) and all-grade cancer detection of 98.0% (48/49).

Conclusion

This feasibility study suggests that the two-step ADC-PSAD-GWR rule may help guide biopsy decisions by identifying a low-risk subgroup where biopsy may be safely deferred. However, due to the retrospective, single-center design and lack of external validation, the findings should be considered hypothesis-generating. No claim of superiority over PI-RADS or reduction in patient-important outcomes (e.g., mortality) is made.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12880-026-02470-5.

Key points

Question: Can a two-step rule combining ADC, PSAD, and gadolinium wash-in/wash-out rate assist in stratifying prostate lesions and guide biopsy decisions within a feasibility setting?

Findings: In this retrospective case series of 110 patients, the two-step rule identified a low-risk subgroup (Q3-low GWR) with zero clinically significant cancers, and showed the potential to defer biopsy in 35% of PI-RADS ≥ 4 benign lesions while detecting all occult csPCa in PI-RADS ≤ 3 lesions.

Clinical relevance: As an exploratory feasibility tool, this objective rule may offer a practical adjunct to mpMRI interpretation by identifying a defined low-risk subgroup for potential biopsy deferral. Prospective external validation is required before clinical implementation.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12880-026-02470-5.

Introduction

Prostate cancer remains a leading cause of cancer-related death in men, with early detection relying heavily on serum prostate-specific antigen (PSA) testing and multiparametric MRI (mpMRI) [1, 2]. However, PSA levels between 4 and 10 ng/mL—the diagnostic grey zone—have limited specificity, driving overdiagnosis and unnecessary biopsies [3, 4]. The introduction of the Prostate Imaging-Reporting and Data System (PI-RADS) v2.1 standardised mpMRI interpretation and is now recommended before biopsy [5, 6]. Yet PI-RADS suffers from moderate specificity and considerable inter-reader variability, leading to false-positive referrals and occult clinically significant prostate cancer (csPCa, ISUP grade group ≥ 2) that escapes detection [6, 7].

Quantitative biomarkers that probe distinct pathophysiological dimensions may complement anatomical scoring. Apparent diffusion coefficient (ADC) from diffusion-weighted imaging inversely reflects tissue cellularity: densely packed malignant glands restrict water diffusion, yielding low ADC values [8–10]. PSA density (PSAD)—serum PSA normalised to prostate volume—integrates tumour burden with gland size and consistently outperforms PSA alone for csPCa prediction [11, 12]. The gadolinium wash-in/wash-out rate (GWR) at 5 min, derived from dynamic contrast-enhanced MRI, quantifies vascular permeability; aggressive tumours exhibit rapid contrast entry and accelerated delayed-phase wash-out due to leaky neovasculature and elevated interstitial pressure [13–17]. In benign prostatic hyperplasia, the microvasculature retains contrast, resulting in low GWR [18, 19]. Preliminary evidence from hepatic and breast imaging indicates that delayed wash-out metrics provide independent diagnostic information [20–22]. However, whether sequential integration of ADC, PSAD, and GWR can correct PI-RADS misclassification in prostate cancer remains unexplored.

We developed a two-step biological framework. First, lesions are assigned to four risk quadrants (Q1–Q4) using cross-validated ADC and PSAD cutoffs. Second, risk is refined by GWR. This study compares the performance of this objective, biology-driven rule against PI-RADS v2.1 for safely avoiding unnecessary biopsies and detecting occult csPCa. The framework uses only standard MRI sequences and simple calculations, offering a practical adjunct to mpMRI interpretation.

Materials and methods

Study design and population

This is a retrospective feasibility study and diagnostic case series. The institutional ethics committee approved this study (approval No. II2026-045-02). Between January 2023 and January 2024, 148 patients with suspected prostate lesions who underwent multiparametric magnetic resonance imaging (mpMRI) were screened. Patients were initially considered for inclusion if they (a) presented with lower urinary tract symptoms or elevated prostate-specific antigen (PSA) and (b) had completed a preoperative mpMRI examination including high-b-value diffusion-weighted imaging (DWI) and dynamic contrast-enhanced (DCE) sequences.

Patients were subsequently excluded according to the following criteria: (1) prior prostate surgery (n = 15); (2) missing clinical, imaging, or pathological data required for quantitative analysis (n = 14); (3) inadequate MRI quality precluding accurate region-of-interest placement (n = 9). The final cohort comprised 110 patients, stratified by postoperative pathology into a clinically significant prostate cancer (csPCa) group (ISUP grade group ≥ 2; n = 42) and a non-csPCa group (n = 68) consisting of benign prostatic hyperplasia (BPH) and low-risk PCa (ISUP 1). Figure 1 shows the patient selection flowchart.

Fig. 1.

Patient selection flowchart and study framework comparing the two-step biological stratification (ADC × PSAD quadrants → GWR refinement) with conventional PI-RADS assessment. From 148 patients with suspected prostate lesions who underwent mpMRI, 38 were excluded (prior surgery, n = 15; missing data, n = 14; inadequate MRI quality, n = 9). The final cohort of 110 patients was stratified by histopathology into non-csPCa (benign or ISUP 1, n = 68) and csPCa (ISUP ≥ 2, n = 42) groups. Both groups were evaluated by the two-step biological rule (ADC-PSAD quadrants → GWR refinement) and by conventional PI-RADS v2.1 scoring to compare diagnostic performance and biopsy avoidance

MRI acquisition

All examinations were performed on a 3.0-T MRI system (Architect, GE Healthcare) with a phased-array abdominal coil, following PI-RADS v2.1 recommendations [6]. The protocol included:

Diffusion-weighted imaging (DWI)

TR 4000–5000 ms, TE minimum, matrix 128 × 128, FOV 32 cm, slice thickness 3 mm, gap 1 mm, b-values 100, 800, 1500 s/mm². Apparent diffusion coefficient (ADC) maps were automatically generated.

Dynamic contrast-enhanced MRI (DCE-MRI)

3D T1-weighted gradient-echo sequence (TR/TE minimum) with temporal resolution of 10 s per phase, 30 consecutive phases, FOV 36 cm, matrix 320 × 192, slice thickness 5 mm (no gap). Gadodiamide (Omniscan^®, GE Healthcare) was administered intravenously at 0.1 mmol/kg and 2.0 mL/s, followed by a 30-mL saline flush.

Calculation of imaging parameters

ADC values were recorded directly from parametric maps. The gadolinium wash-in/wash-out rate (GWR) was calculated as:

where Speak denotes peak signal intensity in the DCE series, S5min signal intensity 5 min after contrast injection, and Spre pre-contrast baseline signal intensity. The 5-minute time point was chosen based on prior evidence [21] and preliminary analysis showing maximal signal difference between PCa and BPH at this delay. For clinical interpretation, a lower GWR value indicates greater contrast retention at 5 min (typically seen in benign or hyperplastic tissue), whereas a higher GWR value reflects accelerated contrast wash-out (commonly associated with malignant neovasculature).

Lesion delineation and ROI analysis

Two genitourinary radiologists (10 and 15 years of experience), blinded to clinical and pathological data, independently analysed images on a GE Advantage Workstation 4.7. For each patient, the index lesion was identified on T2-weighted imaging and high-b-value DWI (b = 1500 s/mm²). Regions of interest (ROIs) were manually drawn along the lesion borders on the slice showing the maximum transverse diameter and then copied to the corresponding ADC map and to the pre-contrast, peak, and 5-minute post-contrast DCE images. Three consecutive slices were averaged. Vessels, urethral lumen, calcifications, necrotic areas, and artefacts were carefully excluded. Interobserver agreement was excellent (intraclass correlation coefficient: ADC 0.92, GWR 0.89). PSA density (PSAD) was calculated as serum PSA divided by prostate volume, with prostate volume measured on T2-weighted images using the ellipsoid formula.

Pathological diagnosis

Two uropathologists (12 and 18 years of experience) reviewed all surgical specimens according to the 2019 International Society of Urological Pathology (ISUP) consensus guidelines. Whole-mount sections were systematically examined for radical prostatectomy specimens; representative tissue blocks were evaluated for transurethral resection specimens. Immunohistochemical staining with basal cell markers (34βE12, p63) and P504S was performed when necessary to confirm the diagnosis. The primary endpoint was clinically significant prostate cancer (csPCa), defined as ISUP grade group ≥ 2 (comparator: benign tissue and ISUP 1). A sensitivity analysis was performed using any-grade prostate cancer (all-grade PCa) as the endpoint, defined as any histopathologically confirmed prostate adenocarcinoma (comparator: benign).

Two-step biological risk stratification framework

The two-step rule integrates three biomarkers that capture distinct pathophysiological dimensions: ADC (cellular density), PSAD (tumour burden), and GWR (vascular permeability). Optimal cutoffs were derived from 5-fold cross-validated receiver operating characteristic (ROC) analysis by maximising the Youden index. To ensure robustness against sampling variability inherent in random dataset splits, the median threshold across all test folds was selected as the final operating point. The test-set median thresholds were ADC 0.747 × 10⁻³ mm²/s, PSAD 0.344 ng/mL², and GWR 9.454% (Table 2).

Table 2.

Cross-validated diagnostic performance and optimal cutoffs of six candidate markers for predicting csPCa

Feature	5-fold cross-validation performance (mean ± SD)
	AUC	Sensitivity	Specificity	PPV	NPV	F1-score
PI-RADS	0.634 ± 0.065	0.492 ± 0.042	0.750 ± 0.094	0.572 ± 0.074	0.693 ± 0.019	0.525 ± 0.027
PSA (ng/mL)	0.627 ± 0.128	0.600 ± 0.114	0.700 ± 0.215	0.617 ± 0.241	0.722 ± 0.097	0.593 ± 0.134
ADC (×10⁻³ mm²/s)	0.658 ± 0.156	0.769 ± 0.224	0.570 ± 0.211	0.549 ± 0.159	0.814 ± 0.188	0.634 ± 0.168
GWR (%)	0.680 ± 0.059	0.908 ± 0.167	0.550 ± 0.190	0.587 ± 0.087	0.939 ± 0.099	0.697 ± 0.029
PSAD (ng/mL²)	0.751 ± 0.084	0.677 ± 0.192	0.830 ± 0.152	0.771 ± 0.181	0.807 ± 0.077	0.691 ± 0.120
TIC type	0.732 ± 0.054	0.646 ± 0.140	0.760 ± 0.065	0.637 ± 0.069	0.773 ± 0.070	0.637 ± 0.089
Optimal cutoffs (Youden index) — probability vs. raw measurement scale
Feature	Probability cutoff (logistic output)			Raw measurement cutoff
	Mean	SD	Median	Mean	SD	Median
PI-RADS	0.49	0.028	0.487	4.501	0.005	4.503
PSA (ng/mL)	0.301	0.042	0.298	15.275	9.11	14.335
ADC (×10⁻³ mm²/s)	0.316	0.128	0.388	0.81	0.12	0.747
GWR (%)	0.268	0.091	0.239	17.879	20.165	9.454
PSAD (ng/mL²)	0.31	0.072	0.305	0.492	0.411	0.344
TIC type	0.43	0.007	0.43	2.517	0.003	2.517

Performance metrics (mean ± SD) were derived from 5-fold stratified cross-validation. Cutoffs were optimised per fold by maximising the Youden index; the median raw measurement across test folds was selected for the two-step rule

Notes: The two-step rule uses the median cutoffs highlighted in bold: ADC 0.747 × 10⁻³ mm²/s, PSAD 0.344 ng/mL², and GWR 9.454%. AUC = area under the receiver operating characteristic curve; PPV = positive predictive value; NPV = negative predictive value

Step 1: ADC-PSAD four-quadrant classification. Patients were assigned to one of four biological quadrants based on ADC and PSAD cutoffs:

• Q1: low ADC / low PSAD.

• Q2: low ADC / high PSAD (malignant-dominant).

• Q3: high ADC / low PSAD (benign-dominant).

• Q4: high ADC / high PSAD.

Step 2: GWR refinement. Within each quadrant, patients were further stratified by GWR into low-risk (GWR < 9.454%) and elevated-risk (GWR ≥ 9.454%) subgroups. The combination defines distinct biological phenotypes.

The clinical impact of the two-step rule was evaluated against conventional PI-RADS v2.1 scoring by comparing the ability to identify csPCa while potentially avoiding unnecessary biopsies. Because this is a feasibility study and lacks external validation, no claim of superiority is made.

Statistical analysis

Statistical analyses were performed using R version 4.5.1 (R Foundation for Statistical Computing, Vienna, Austria) and SPSS version 26.0 (IBM Corp., Armonk, NY, USA). Normality of continuous variables was assessed with the Kolmogorov-Smirnov test. Normally distributed data are presented as mean ± standard deviation and compared with the independent two-sample t-test; non-normally distributed data are presented as median (interquartile range) and compared with the Mann-Whitney U test. Categorical variables are reported as frequencies (percentages) and analysed with Pearson’s chi-square test or Fisher’s exact test.

Spearman’s rank correlation coefficients (ρ) evaluated relationships among imaging biomarkers, clinical parameters, and csPCa status. Diagnostic performance of individual markers was assessed with ROC analysis; the area under the curve (AUC) with 95% confidence intervals was computed via the DeLong method. Optimal cutoffs were determined by maximising the Youden index. Five-fold stratified cross-validation (7:3 split) was performed for six candidate markers to assess out-of-sample generalisability, reporting mean AUC, sensitivity, specificity, positive predictive value, negative predictive value, and F1-score across test folds. Pairwise AUC comparisons used the DeLong test.

Calibration analysis was performed by plotting the predicted probability of csPCa against the observed proportion of csPCa using a calibration curve. The slope and intercept of the calibration curve were computed to assess goodness-of-fit. The Hosmer-Lemeshow test was also performed to evaluate calibration; a non-significant p-value indicates good calibration.

Mediation analysis with 5,000 bootstrap resamples examined the bidirectional relationship between ADC and PSAD and tested the independence of GWR from PSAD. Direct and indirect effects with 95% confidence intervals were estimated, adjusting for age and body mass index. Hierarchical forward logistic regression (likelihood ratio) evaluated the incremental predictive value of ADC, GWR, and PSAD beyond conventional imaging scores (PI-RADS and time-intensity curve type). Model fit was assessed with Nagelkerke’s R² and the Hosmer-Lemeshow test. For the two-step stratification framework, cross-tabulation determined csPCa and overall PCa rates within each ADC-PSAD quadrant and GWR subgroup, with Wilson 95% confidence intervals. Decision curve analysis quantified the net clinical benefit of each marker and the combined rule across a range of threshold probabilities. All tests were two-sided; p < 0.05 was considered statistically significant.

Note on external validation: Because this is a single-center retrospective feasibility study, no independent external validation cohort was available. The cross-validation results should be interpreted as internal performance estimates that may not generalise to other settings. Future multi-center validation is required.

Results

Study population and baseline characteristics

The final cohort comprised 110 patients (mean age, 70 ± 9 years; all men), including 42 with csPCa (ISUP grade group ≥ 2) and 68 non-csPCa controls (benign prostatic hyperplasia and low-risk PCa [ISUP 1]). Baseline characteristics are summarised in Table 1. Relative to non-csPCa controls, the csPCa group had lower ADC values (0.70 ± 0.20 × 10⁻³ mm²/s vs. 0.80 ± 0.20 × 10⁻³ mm²/s, p < 0.001), higher PSAD (median 0.8 ng/mL² vs. 0.1 ng/mL², p < 0.001), and higher GWR (median 56.4% vs. 19.8%, p < 0.001). PI-RADS scores and TIC types also differed between groups (p < 0.01 and p < 0.001, respectively). Age, body mass index, height, and weight did not differ (Table 1). Spearman correlation analysis (Fig. 2, Fig. S1, Table S1) showed a strong inverse correlation between ADC and PSAD (ρ = -0.52, p < 0.001). GWR correlated moderately with csPCa (ρ = 0.35, p < 0.001) and weakly with ADC (ρ = -0.28, p = 0.003) and PSAD (ρ = 0.23, p = 0.017). Box-and-whisker plots (Fig. 3) illustrate the separation between groups for ADC, PSAD, and GWR.

Table 1.

Baseline demographic, clinical, and imaging characteristics of the study cohort stratified by csPCa status

Variable	All patients (N = 110)	Non-csPCa (n = 68)	csPCa (n = 42)	Test statistica	p-valuea
		(BPH + low-risk PCa)	(clinically significant PCa)
ADC (×10⁻³ mm²/s)	0.78 ± 0.22	0.80 ± 0.20	0.70 ± 0.20	t = 4.329	< 0.001
PSA (ng/mL)	10.9 (3.2, 25.4)b	7.9 (3.2, 12.6)b	22.3 (7.5, 54.3)b	U = 827.5	< 0.001
GWR (%)	29.3 (-5.3, 63.9)b	19.8 (-5.5, 45.1)b	56.4 (28.0, 84.9)b	U = 837.5	< 0.001
PSAD (ng/mL²)	0.2 (0.0, 0.6)b	0.1 (0.0, 0.3)b	0.8 (0.2, 1.5)b	U = 575.5	< 0.001
Volume (mm³)	43383.9 (28589.0, 58158.8)b	52104.5 (36254.1, 67954.9)b	35282.5 (24471.3, 46093.7)b	U = 837.5	< 0.001
TTP (seconds)	71.0 (43.8, 98.3)b	90.0 (51.3, 128.8)b	59.5 (44.1, 74.9)b	U = 848.5	< 0.001
Height (m)	1.7 (1.6, 1.8)b	1.7 (1.6, 1.8)b	1.7 (1.6, 1.8)b	U = 1245.0	0.598
Weight (kg)	65.3 ± 9.1	65.4 ± 9.0	65.0 ± 9.4	t = 0.236	0.814
BMI (kg/m²)	23.4 ± 2.9	23.5 ± 3.0	23.3 ± 2.8	t = 0.247	0.805
Age (years)	69.5 ± 8.7	68.6 ± 9.1	70.9 ± 8.1	t = -1.310	0.193
TIC type, n (%)	Type 1: 24 (21.8%)	Type 1: 21 (30.9%)	Type 1: 3 (7.1%)	χ² = 13.809, df = 2	< 0.01
	Type 2: 40 (36.4%)	Type 2: 27 (39.7%)	Type 2: 13 (31.0%)
	Type 3: 46 (41.8%)	Type 3: 20 (29.4%)	Type 3: 26 (61.9%)
PI-RADS score, n (%)	2: 19 (17.3%)	2: 15 (22.1%)	2: 4 (9.5%)	χ² = 14.540, df = 4	0.006**
	3: 24 (21.8%)	3: 16 (23.5%)	3: 8 (19.0%)
	4: 34 (30.9%)	4: 25 (36.8%)	4: 9 (21.4%)
	5: 32 (29.1%)	5: 12 (17.6%)	5: 20 (47.6%)
	6: 1 (0.9%)	6: 0 (0.0%)	6: 1 (2.4%)
Prostate cancer (PCa), n (%)	Benign: 61 (55.5%)	Benign: 61 (89.7%)	Benign: 0 (0.0%)	χ² = 80.987, df = 1	< 0.001
	PCa: 49 (44.5%)	PCa: 7 (10.3%)	PCa: 42 (100.0%)

Data are presented as mean ± standard deviation, median (interquartile range), or frequency (%)

Notes: a Group comparisons were performed using the independent t-test for normally distributed variables, the Mann–Whitney U test for non-normally distributed variables, and the χ² test for categorical variables. b Quartiles were derived under the assumption of symmetric distribution. csPCa = clinically significant prostate cancer (ISUP ≥ 2); PSAD = PSA density; GWR = gadolinium wash-in/wash-out rate; TIC = time–intensity curve; PI-RADS = Prostate Imaging–Reporting and Data System v2.1

Fig. 2.

Spearman correlation network of ADC, PSAD, and GWR with clinically significant prostate cancer (csPCa) — mechanistic basis for the two-step stratification rule. Circular network visualisation of Spearman’s rank correlation coefficients (ρ). The strong inverse correlation between ADC and PSAD (ρ = −0.52, p < 0.001) supports their paired use in the first-step quadrants. GWR correlates moderately with csPCa (ρ = 0.35) but weakly with ADC (ρ = −0.28) and PSAD (ρ = 0.23), confirming its independent vascular information as a second-step modifier

Fig. 3.

Comparative distribution of core biological and imaging biomarkers between patients with clinically significant prostate cancer (csPCa) and non-csPCa controls. Box-and-whisker plots with overlaid jittered scatter points for (A) ADC, (B) GWR, (C) PI-RADS score, (D) PSA, (E) PSAD, and (F) TIC type. Box boundaries indicate first and third quartiles; the central line denotes the median. Non-csPCa group (n = 68) includes benign and ISUP 1 cases; csPCa group (n = 42) includes ISUP ≥ 2 cancers. Group comparisons used the Mann–Whitney U test or independent t-test as appropriate. ***p < 0.001, **p < 0.01, *p < 0.05; ns = not significant

Single-marker diagnostic performance and cross-validation

ROC analysis on the full cohort (Table S2) identified PSAD as the strongest single discriminator of csPCa (AUC = 0.798; 95% CI: 0.711–0.886), with sensitivity 71.4% and specificity 82.4% at the Youden-derived cutoff of 0.312 ng/mL². GWR showed the highest sensitivity (92.9%) at a cutoff of 9.225%. ADC achieved an AUC of 0.727. In 5-fold stratified cross-validation (Table 2), PSAD maintained the highest mean AUC (0.751 ± 0.084) with specificity 83.0%, whereas GWR had sensitivity 90.8% and negative predictive value 93.9%. PI-RADS yielded a cross-validated AUC of 0.634 ± 0.065, similar to conventional PSA (AUC = 0.627 ± 0.128). Pairwise DeLong tests (Fig. S2) confirmed that PSAD outperformed PI-RADS (p < 0.05). Decision curve analysis (Fig. 4) showed that PSAD provided the greatest net benefit, and GWR added benefit in intermediate-risk zones.

Fig. 4.

Cross-validated diagnostic metrics, confusion matrices, and decision curve analysis of the six biomarkers, illustrating the rationale for the two-step ADC-PSAD-GWR stratification framework. For each marker, 5-fold cross-validated ROC curves, confusion matrices, and decision curve analyses (DCA) are shown. (A) ADC provides high sensitivity and NPV. (B) PSAD achieves the highest AUC (0.751) and greatest net benefit. (C) GWR shows the highest sensitivity (92.9%) and complements intermediate-risk stratification. (D) PSA, (E) PI-RADS, and (F) TIC type exhibit lower or narrower net benefit. The complementary patterns of ADC, PSAD, and GWR underpin the sequential two-step rule

Mechanistic validation: mediation analysis and incremental value

Mediation analysis with 5,000 bootstrap resamples (Table 3) demonstrated bidirectional partial mediation between ADC and PSAD. PSAD mediated 30.6% of the effect of ADC on csPCa (indirect effect = -0.231, p = 0.020), and ADC mediated 25.9% of the effect of PSAD (indirect effect = 0.024, p = 0.026). Direct effects remained significant in both directions. GWR showed no mediation via PSAD (indirect effect = 0, p = 0.754) and retained a direct effect on csPCa (p = 0.012). Hierarchical logistic regression (Table S3) showed that ADC, GWR, and PSAD each contributed incremental variance beyond PI-RADS and TIC type (ΔR² = 0.037, 0.036, and 0.041, respectively; all p < 0.05). PI-RADS became non-significant after ADC entered the model. The full model explained 32.9% of the variance in csPCa (Nagelkerke R²).

Table 3.

Mediation analysis of the relationships among ADC, PSAD, GWR, and csPCa

Pathway	Effect type	Effect	95% CI	SE	z / t	p	Proportion
		estimate
Part 1. Bidirectional ADC → PSAD mediation (supports four-quadrant framework)
ADC → PSAD → csPCa	Indirect (a×b)	-0.231	-0.464 to -0.085	0.1	-2.318	0.02	30.60%
ADC → PSAD (a)	X → M	-3.622**	-5.240 to -2.003	0.816	-4.439	< 0.001	—
PSAD → csPCa (b)	M → Y	0.064**	0.017 to 0.110	0.023	2.728	0.008	—
ADC → csPCa (c’)	Direct	-0.522*	-0.939 to -0.105	0.21	-2.484	0.015	—
ADC → csPCa (c)	Total	-0.753**	-1.146 to -0.359	0.198	-3.795	< 0.001	—
PSAD → ADC → csPCa	Indirect (a×b)	0.024	0.004 to 0.047	0.011	2.231	0.026	25.90%
PSAD → ADC (a)	X → M	-0.047**	-0.067 to -0.027	0.01	-4.621	< 0.001	—
ADC → csPCa (b)	M → Y	-0.522*	-0.939 to -0.105	0.21	-2.484	0.015	—
PSAD → csPCa (c’)	Direct	0.064**	0.017 to 0.110	0.023	2.728	0.008	—
PSAD → csPCa (c)	Total	0.094**	0.050 to 0.139	0.023	4.174	< 0.001	—
Part 2. GWR pathways — no significant mediation via PSAD (supports second-step independence)
GWR → PSAD → csPCa	Indirect (a×b)	0	-0.000 to 0.001	0	0.314	0.754	0%
GWR → PSAD (a)	X → M	0.001	-0.008 to 0.011	0.005	0.271	0.787	—
PSAD → csPCa (b)	M → Y	0.064**	0.017 to 0.110	0.023	2.728	0.008	—
GWR → csPCa (c’)	Direct	0.003*	0.001 to 0.005	0.001	2.57	0.012	—
GWR → csPCa (c)	Total	0.003*	0.001 to 0.005	0.001	2.564	0.012	—
PSAD → GWR → csPCa	Indirect (a×b)	0.006	-0.002 to 0.022	0.006	1.023	0.306	0%
PSAD → GWR (a)	X → M	2.169	-1.615 to 5.952	1.908	1.137	0.258	—
GWR → csPCa (b)	M → Y	0.003*	0.001 to 0.005	0.001	2.57	0.012	—
PSAD → csPCa (c’)	Direct	0.064**	0.017 to 0.110	0.023	2.728	0.008	—
PSAD → csPCa (c)	Total	0.094**	0.050 to 0.139	0.023	4.174	< 0.001	—

Analyses were adjusted for age and body mass index (n = 110). Indirect effects and 95% confidence intervals were estimated with 5,000 bootstrap resamples

Notes: Part 1 shows bidirectional partial mediation between ADC and PSAD (each mediates ≈ 26–31% of the other’s effect on csPCa), supporting their paired use in the first-step quadrants. Part 2 demonstrates that GWR exerts a direct effect on csPCa independently of PSAD (indirect effect p = 0.754), justifying its application as a second-step vascular modifier. *p < 0.05, **p < 0.01

Model calibration assessment

To rigorously evaluate the predictive performance of the combined ADC, GWR, and PSAD model, we performed a calibration analysis (Fig. S3; Table S5). The calibration curve (linear fit) demonstrated excellent agreement between the predicted probability of csPCa and the observed proportion in the cohort. The calibration intercept was − 0.013 (ideal = 0) and the calibration slope was 1.034 (ideal = 1), indicating minimal over- or under-estimation across the range of predicted risks. The Hosmer-Lemeshow test was not statistically significant (χ² = 6.099, df = 8, p = 0.636), further supporting acceptable model calibration. The model exhibited good discriminatory capacity, with an area under the ROC curve (AUC) of 0.788 (95% CI: 0.699–0.877). These findings directly address the reviewer’s request for calibration analysis beyond the Hosmer-Lemeshow test alone.

Two-step framework: quadrant distribution and risk stratification

Using the cross-validated median cutoffs (ADC 0.747 × 10⁻³ mm²/s, PSAD 0.344 ng/mL², GWR 9.454%), patients were assigned to four ADC-PSAD quadrants and further stratified by GWR (Fig. 5A, C; Table 4).

Fig. 5.

Two-step biological risk stratification framework integrating ADC, PSAD, and GWR.Cutoffs (test-set medians): ADC 0.747 × 10⁻³ mm²/s, PSAD 0.344 ng/mL², GWR 9.454%. (A, B) Primary analysis: clinically significant prostate cancer (csPCa, ISUP ≥ 2 vs. benign/ISUP 1). (A) First-step ADC-PSAD quadrants partition the cohort into four risk phenotypes. Q2 (low ADC/high PSAD) is the malignant-dominant quadrant (77.4% csPCa); Q3 (high ADC/low PSAD) is benign-dominant (10.9% csPCa). (B) Second-step refinement by GWR. Within Q3, the low-GWR subgroup contains zero csPCa (0/20), defining a safe biopsy-avoidance zone. High GWR substantially elevates risk in intermediate quadrants Q1 and Q4. (C, D) Sensitivity analysis: all-grade prostate cancer (PCa, any ISUP grade vs. benign). (C) The same ADC-PSAD cutoffs produce analogous quadrant phenotypes. (D) GWR refinement yields consistent risk patterns; notably, Q2-high GWR reaches 100% cancer prevalence (29/29, including 7 ISUP 1 cases). The two-step framework demonstrates robust performance across both stringent and inclusive cancer definitions, correcting PI-RADS misclassification and reducing unnecessary biopsies by 35–43%

Table 4.

Two-step risk stratification by ADC-PSAD quadrants and GWR for the primary endpoint (csPCa) and sensitivity analysis (all-grade PCa)

Quadrant	Definition	GWR group	Total (n)	csPCa (ISUP ≥ 2)			PCa (all grades)
				n	% of quadrant	Rate, % (95% CI)†	n	% of quadrant	Rate, % (95% CI)†
Q1	Low ADC & Low PSAD	Low	9	1	11.10%	11.1 (1.4–43.9)	1	11.10%	11.1 (1.4–43.9)
Low ADC / Low PSAD		High	13	6	46.20%	46.2 (22.2–71.6)	6	46.20%	46.2 (22.2–71.6)
Q2	Low ADC & High PSAD	Low	2	2	100%	100.0 (22.4–100.0)	2	100%	100.0 (22.4–100.0)
Low ADC / High PSAD		High	29	22	75.90%	75.9 (57.9–88.4)	29	100%	100.0 (88.1–100.0)
Q3	High ADC & Low PSAD	Low	20	0	0%	0.0 (0.0–16.8)	0	0%	0.0 (0.0–16.8)
High ADC / Low PSAD		High	26	5	19.20%	19.2 (7.9–38.2)	5	19.20%	19.2 (7.9–38.2)
Q4	High ADC & High PSAD	Low	3	0	0%	0.0 (0.0–70.8)	0	0%	0.0 (0.0–70.8)
High ADC / High PSAD		High	8	6	75.00%	75.0 (39.7–94.7)	6	75.00%	75.0 (39.7–94.7)
Total (csPCa cohort)				42	38.20%	38.2 (29.5–47.7)	—	—	—
Total (overall PCa cohort)				—	—	—	49	44.50%	44.5 (35.4–54.0)

Cutoffs (test-set medians): ADC 0.747 × 10⁻³ mm²/s, PSAD 0.344 ng/mL², GWR 9.454%. Data are n or % (95% Wilson confidence interval)

Notes: Primary analysis defines csPCa as ISUP ≥ 2 (comparator: benign + ISUP 1); sensitivity analysis defines PCa (all grades) as any histopathologically confirmed prostate cancer (comparator: benign). † In Q2-high GWR, csPCa rate is 75.9% (22/29) whereas all-grade PCa rate reaches 100% (29/29); the seven additional cases are ISUP 1 cancers. Q3-low GWR (grey-shaded in workflow) identifies a biopsy-avoidance subgroup with zero cancers in both analyses (0/20)

• Q2 (low ADC/high PSAD) contained 31 patients and represented the malignant-dominant phenotype. In the primary analysis (csPCa: ISUP ≥ 2 vs. benign/ISUP 1), csPCa rates were 100% (2/2) in the low-GWR subgroup and 75.9% (22/29) in the high-GWR subgroup (Fig. 5B). In the sensitivity analysis (all-grade PCa: any cancer vs. benign), the Q2-high GWR subgroup harbored cancer in all 29 patients (100%; 95% CI: 88.1–100.0) (Fig. 5D).

• Q3 (high ADC/low PSAD) contained 46 patients and was predominantly benign. The Q3-low GWR subgroup (n = 20) had no csPCa cases (0/20; 95% CI: 0-16.8%) and no all-grade PCa cases, defining a potential biopsy-avoidance subgroup.

• Q1 and Q4 were intermediate quadrants. In Q1 (n = 22), csPCa rates increased from 11.1% (1/9) in the low-GWR subgroup to 46.2% (6/13) in the high-GWR subgroup. In Q4 (n = 11), csPCa was absent in the low-GWR subgroup (0/3) but present in 75.0% (6/8) of the high-GWR subgroup.

GWR thus reclassified risk substantially in these intermediate quadrants. Findings for all-grade PCa were consistent with the primary analysis (Table 4).

Direct comparison with PI-RADS: correction of misclassification

The clinical impact of the two-step rule was evaluated against PI-RADS v2.1 scoring (Table 5).

Table 5.

Correction of PI-RADS misclassification by the two-step ADC-PSAD-GWR rule: primary analysis (csPCa) and sensitivity analysis (all-grade PCa)

Quadrant	GWR group	PI-RADS ≤ 3 (scores 2–3)				PI-RADS ≥ 4 (scores 4–5)
		N	csPCa, n (%)	PCa, n (%)	Decision (Benefit)	N	csPCa, n (%)	PCa, n (%)	Decision (Benefit)
Q1	Low	2	0 (0.0%)	0 (0.0%)	Avoid (Safe)	7	1 (14.3%)	1 (14.3%)	Avoid (Missed 1 cancer)
Low ADC / Low PSAD	High	5	0 (0.0%)	0 (0.0%)	Recommend	8	6 (75.0%)	6 (75.0%)	Recommend
Q2	Low	1	1 (100%)	1 (100%)	Recommend (Detected occult)	1	1 (100%)	1 (100%)	Recommend
Low ADC / High PSAD	High	4	3 (75.0%)	4 (100.0%)	Recommend (Detected occult)	25	19 (76.0%)	25 (100.0%)†	Recommend
Q3	Low	13	0 (0.0%)	0 (0.0%)	Avoid (Safe)	7	0 (0.0%)	0 (0.0%)	Avoid (Avoided unnecessary bx)
High ADC / Low PSAD	High	12	5 (41.7%)	5 (41.7%)	Recommend (Detected occult)	14	0 (0.0%)	0 (0.0%)	Recommend
Q4	Low	1	0 (0.0%)	0 (0.0%)	Recommend	2	0 (0.0%)	0 (0.0%)	Recommend
High ADC / High PSAD	High	2	2 (100%)	2 (100%)	Recommend (Detected occult)	6	4 (66.7%)	4 (66.7%)	Recommend
Total (PI-RADS ≤ 3)		43	12 (27.9%)	12 (27.9%)	Avoid: 17 / Rec: 26	Occult cancer detected: 12/12 (100%) for both csPCa and all-grade PCa
Total (PI-RADS ≥ 4)						67	30 (44.8%)	39 (58.2%)	Avoid: 13 / Rec: 54
									Unnecessary bx avoided: 13/37 (35.1%) for csPCa; 13/28(46.4%) for all-grade PCa

Notes: † In the PI-RADS ≥ 4 / Q2-high GWR subgroup, all 25 patients had all-grade PCa (100%); the remaining two patients in Q2-high GWR were in the PI-RADS ≤ 3 category (both PCa). One csPCa (which was also an all-grade PCa) was missed by the rule in the PI-RADS ≥ 4 / Q1-low GWR subgroup. Clinical benefit is summarized in parentheses within the Decision column. The rule detected 100% of occult cancers in PI-RADS ≤ 3 and avoided 35.1% (csPCa) to 46.4% (all-grade PCa) of unnecessary biopsies in PI-RADS ≥ 4

PI-RADS ≤ 3 lesions. Among 43 patients with PI-RADS ≤ 3, 12 occult csPCa were identified by histopathology. The two-step rule recommended biopsy in 26 of these patients and successfully detected all 12 occult csPCa (100%) in the primary analysis. Sensitivity analysis confirmed identical detection for all-grade PCa (12/12).

PI-RADS ≥ 4 lesions. Among 67 patients with PI-RADS ≥ 4, 30 had csPCa and 37 were benign or low-grade (ISUP 1) in the primary analysis. The two-step rule avoided biopsy in 13 patients, all of whom were benign. This corresponds to a 35.1% (13/37) reduction in unnecessary biopsies for the primary endpoint. In the sensitivity analysis, where the 7 ISUP 1 cancers are classified as positive events, there were 39 all-grade PCa and 28 benign lesions among PI-RADS ≥ 4 patients. The two-step rule avoided biopsy in 13 benign cases, resulting in an unnecessary biopsy avoidance rate of 46.4% (13/28).

Overall performance. Across the entire cohort, the two-step rule achieved a csPCa detection rate of 97.6% (41/42) and a safe avoidance rate of 96.6% (28/29) in the primary analysis. Sensitivity analysis demonstrated an all-grade PCa detection rate of 98.0% (48/49) with an identical safe avoidance rate of 96.6%.

One csPCa (which was also an all-grade PCa) was missed by the rule in the PI-RADS ≥ 4 / Q1-low GWR subgroup.

Illustrative cases

Figure 6 shows representative cases. The left panels (A–F) depict a patient with PI-RADS 4, ADC 1.026 × 10⁻³ mm²/s, PSAD 0.226 ng/mL², and GWR 3.19%. This profile corresponds to Q3-low GWR, the nominally ‘biopsy-avoidance’ phenotype. Histopathology confirmed benign prostatic hyperplasia. The right panels (a–f) show a patient with PI-RADS 5, ADC 0.599 × 10⁻³ mm²/s, PSAD 1.209 ng/mL², and GWR 79.78% — Q2-high GWR, the malignant-dominant phenotype. Targeted biopsy revealed ISUP 1 adenocarcinoma (overall PCa positive, csPCa negative). The two-step rule correctly mandated biopsy, ensuring low-grade cancer was sampled.

Fig. 6.

Two-step ADC-PSAD-GWR stratification corrects PI-RADS misclassification: benign avoidance in PI-RADS 4 and occult cancer detection in high-grade imaging. Left panels (A–F): PI-RADS 4 lesion with high ADC (1.026 × 10⁻³ mm²/s), low PSAD (0.226 ng/mL²), and low GWR (3.19%) → Q3-low GWR. The two-step rule recommended biopsy avoidance; histopathology confirmed benign hyperplasia. Right panels (a–f): PI-RADS 5 lesion with low ADC (0.599 × 10⁻³ mm²/s), high PSAD (1.209 ng/mL²), and high GWR (79.78%) → Q2-high GWR. The rule correctly mandated biopsy; pathology revealed ISUP 1 adenocarcinoma (all-grade PCa positive, csPCa negative). The biological framework objectively reclassifies ambiguous imaging findings, avoiding unnecessary procedures without missing cancer

Discussion

This study presents a two-step framework that integrates apparent diffusion coefficient (ADC), PSA density (PSAD), and gadolinium wash-in/wash-out rate (GWR) to stratify prostate lesions. The rule identifies a malignant-dominant quadrant requiring biopsy and a biopsy-avoidance subgroup that, in our cohort, safely excluded clinically significant prostate cancer (csPCa). Compared with PI-RADS v2.1 within this retrospective single-center cohort, the framework demonstrated the potential to reduce unnecessary biopsies and detect occult csPCa. The following discussion addresses the biological and statistical foundations of the rule, compares it with existing stratification tools, and outlines its clinical implications, limitations, and future directions.

Biological rationale, orthogonality, and threshold considerations

Figure 7 illustrates the distinct pathophysiological dimensions captured by the three biomarkers. In benign or hyperplastic tissue, preserved glandular architecture permits free water diffusion (high ADC), intact microvasculature retains contrast (low GWR), and PSA remains largely confined to the ductal system (low PSAD). In clinically significant cancer, increased cellularity restricts diffusion (low ADC), disordered angiogenesis accelerates contrast wash-out (high GWR), and basal membrane disruption releases PSA into the circulation (high PSAD). These parameters exhibit only weak-to-moderate inter-correlations (|ρ| ≤ 0.52; Table S1), confirming they probe orthogonal biological axes.

Fig. 7.

Pathophysiological basis of the two-step ADC-PSAD-GWR stratification framework. Left panel (ADC): Cellular density and water diffusion. In benign or hyperplastic glands, preserved luminal architecture and organised stroma permit unrestricted Brownian motion of water molecules, yielding high apparent diffusion coefficient (ADC) values. In clinically significant prostate cancer, densely packed malignant cells, nuclear enlargement, and expanded nucleoli restrict both extracellular and intracellular water mobility, resulting in markedly reduced ADC. Middle panel (GWR): Microvascular permeability and contrast wash-out. Benign tissue maintains relatively intact capillary networks with functional endothelial barriers; following contrast administration, gadolinium chelates remain largely confined to the interstitial compartment, producing low gadolinium wash-in/wash-out rates (GWR). In malignancy, disordered angiogenesis generates tortuous, leaky vessels with disrupted pericyte coverage. Combined with elevated interstitial fluid pressure, this facilitates rapid contrast entry during the arterial phase and accelerated delayed-phase wash-out, manifesting as high GWR. Right panel (PSA): Volume-adjusted tumour burden. In benign prostatic tissue, prostate-specific antigen (PSA) is secreted by glandular epithelial cells into the ductal lumina and is largely retained by an intact basal cell layer and basement membrane, limiting its entry into the systemic circulation. Consequently, serum PSA density (PSAD)—PSA normalised to prostate volume—remains low. In clinically significant cancer, disruption of the basal membrane, loss of normal glandular polarity, and increased microvascular invasion permit PSA to gain direct access to the systemic circulation and lymphatic drainage. This results in elevated serum PSA disproportionate to gland volume, reflected in high PSAD values. The three parameters exhibit only weak-to-moderate inter-correlations (|ρ| ≤ 0.52), indicating that they capture orthogonal information regarding cellularity, vascular dynamics, and volume-adjusted secretory burden, respectively

This orthogonality underpins the two-step design. ADC and PSAD were inversely correlated, and mediation analysis confirmed bidirectional partial mediation: each transmitted approximately one-quarter to one-third of the other’s effect on csPCa. This coupling between cellular density and volume-adjusted secretory burden supports their paired use in the first-step quadrants. GWR, in contrast, exhibited no significant mediation via PSAD and retained an independent direct effect. Its weak correlations with ADC and PSAD confirm that GWR probes vascular permeability separately, justifying its sequential application as a second-step modifier that refines risk when cellularity and tumour burden are inconclusive [14–17].

The derived ADC cutoff (0.747 × 10⁻³ mm²/s) lies within the range recommended by PI-RADS v2.1 for the peripheral zone (0.75–0.90 × 10⁻³ mm²/s) [6], supporting its internal consistency. The PSAD threshold (0.344 ng/mL²) falls between conservative biopsy-avoidance values (0.10–0.15 ng/mL²) [12] and higher surgical cutoffs, reflecting the intended balance between sensitivity and specificity. The GWR threshold (9.454%) is consistent with evidence that delayed wash-out discriminates malignant from benign tissue across multiple organs [20–22]. Of note, these thresholds are protocol- and scanner-specific; local recalibration is strongly recommended before clinical application in other settings.

Comparison with existing approaches and distinction from qualitative PI-RADS

PI-RADS plus PSAD models. Combining PI-RADS with a PSAD cutoff of 0.15 ng/mL² improves csPCa prediction [12]. The present framework extends this strategy by incorporating ADC as a direct cellularity metric and GWR as a vascular modifier, yielding a fully quantitative rule that does not rely on subjective PI-RADS categorisation.

Multiparametric scoring systems (MSI) and serum-based panels (SER). Nomograms such as the MRI-based Stanford MSI and serum tests like 4Kscore or SelectMDX reduce unnecessary biopsies by 30–40% but require additional testing or proprietary algorithms [23, 24]. The two-step rule achieves comparable reduction using only data acquired during routine mpMRI, with no incremental cost or specimen collection. Moreover, the transparency of the rule—each component corresponds to a defined biological process—contrasts with the “black box” character of many commercial panels.

Radiomics and machine learning. Radiomic signatures and deep learning models have shown promise for csPCa detection but often lack interpretability and are sensitive to scanner and protocol variability [25–27]. The ADC-PSAD-GWR framework is transparent, relying on standardised measurements with established pathophysiological correlates. This interpretability facilitates external validation and clinical adoption.

Advanced diffusion and perfusion techniques. Intravoxel incoherent motion, diffusion kurtosis imaging, and quantitative DCE parameters (Ktrans, kep) improve specificity but demand specialised sequences and offline post-processing [28, 29]. The dimensionless GWR formula used here normalises inter-scan variability and can be measured on any standard workstation, offering practical advantages over complex pharmacokinetic models.

Fundamental distinction from PI-RADS. PI-RADS v2.1 assigns a semi-quantitative score based on visual assessment, with moderate specificity and inter-reader variability [6]. The two-step rule complements this approach by replacing subjective interpretation with objective, continuous measurements: ADC quantifies cellular density, PSAD corrects PSA for gland volume, and GWR captures vascular wash-out. These measurements reduce reader dependence and explicitly separate cellular and vascular information. This separation enables granular risk refinement; for instance, a PI-RADS 4 lesion with high ADC, low PSAD, and slow wash-out is reclassified as low risk, avoiding unnecessary biopsy.

It is important to emphasize that this two-step framework is not intended to replace PI-RADS entirely but rather to serve as an objective adjunct [30]. In cases with unequivocal PI-RADS 5 findings, the clinical decision to biopsy is already well-established, and the model serves primarily as confirmatory support. However, in the diagnostic grey zones of PI-RADS 3 and the heterogeneous PI-RADS 4 category—where overdiagnosis and inter-reader variability are most problematic—the objective biological data provided by ADC, PSAD, and GWR offer critical, actionable refinement to guide biopsy decisions more confidently.

Internal statistical rigor and calibration

Five-fold cross-validation, mediation analysis, and hierarchical logistic regression were performed as detailed in the Methods. Model calibration, assessed via calibration curve (intercept − 0.013, slope 1.034) and Hosmer-Lemeshow test (p = 0.636), confirmed good agreement between predicted and observed csPCa probabilities. Decision curve analysis further supported the net clinical benefit of the combined rule. In terms of clinical integration, the rule can be applied using standard MRI data; for PI-RADS 3 or ambiguous PI-RADS 4 lesions, objective ADC, PSAD, and GWR measurements offer actionable refinement without requiring additional cost or specialised sequences.

Limitations and future directions

Key limitations include the retrospective single-centre design, absence of external validation, use of surrogate endpoints (ISUP grade rather than patient-important outcomes), and manual ROI placement. The 5-minute GWR time point is protocol-specific. These factors render the current findings hypothesis-generating. Future work must prioritise prospective multi-centre validation with hard clinical endpoints, and the development of automated segmentation and calculation tools to facilitate clinical translation. The Q3-low GWR phenotype warrants evaluation in active surveillance cohorts.

Conclusion

The two-step ADC-PSAD-GWR rule defines biologically distinct risk phenotypes and, in this retrospective feasibility cohort, demonstrated the potential to correct PI-RADS misclassifications by identifying a low-risk subgroup for potential biopsy deferral. By sequentially integrating quantitative measures of cellular density, volume-adjusted tumour burden, and vascular permeability, this objective framework offers a practical and interpretable adjunct to mpMRI interpretation. However, given the single-centre, retrospective design and the lack of external validation, these findings are best considered hypothesis-generating. Future multi-centre validation with patient-important outcomes is essential before any definitive claims of clinical utility can be made.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

H. C. and T. C. (co-first authors) wrote the main manuscript text and prepared figures. G. L., X. T., M. H., and W. L. performed the experiments and analyzed the data. Q. X. and Z. M. (co-corresponding authors) supervised the project and revised the manuscript. All authors reviewed the manuscript.

Funding

This work was supported by the Hospital National Natural Science Foundation Cultivation Project (Grant No. 2021GZRPYM06) and the Five-Five Project of the Third Affiliated Hospital of Sun Yat-sen University (Grant No. 2023WW605). The funders had no role in study design, data collection, analysis, decision to publish, or preparation of the manuscript.

Data availability

The datasets generated and/or analysed during the current study are not publicly available due to patient privacy and institutional restrictions, but are available from the corresponding author on reasonable request.

Declarations

Ethics approval

This study was performed in accordance with the ethical standards of the Declaration of Helsinki and its later amendments. The study protocol was approved by the Institutional Review Board (IRB) of the Third Affiliated Hospital of Sun Yat-sen University (Approval No. II2026-045-02). All procedures involving human participants were conducted in compliance with the approved protocol.

Human Ethics

This study was conducted in accordance with the Declaration of Helsinki. Ethical approval was obtained from the Institutional Review Board of the Third Affiliated Hospital of Sun Yat-sen University (Approval No. II2026-045-02), which also granted a waiver of informed consent due to the retrospective design. The use of human data complied with all applicable national and institutional guidelines.

Consent to participate

Due to the retrospective nature of this study and the use of anonymized clinical and imaging data, the Institutional Review Board approved a waiver of written informed consent. All patient data were handled in accordance with relevant confidentiality regulations.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.