Chan Woo Wee and BumSup Jang contributed equally to this work.
This study aimed to develop a model for predicting pathologic extracapsular extension (ECE) and seminal vesicle invasion (SVI) while integrating magnetic resonance imagingbased Tstaging (cT_{MRI}, cT1ccT3b).
A total of 1,915 who underwent radical prostatectomy between 2006–2016 met the inclusion/exclusion criteria. We performed a multivariate logistic regression analysis as well as Bayesian network (BN) modeling based on possible confounding factors. The BN model was internally validated using 5fold validation.
According to the multivariate logistic regression analysis, initial prostatespecific antigen (iPSA) (β=0.050, p < 0.001), percentage of positive biopsy cores (PPC) (β=0.033, p < 0.001), both lobe involvement on biopsy (β=0.359, p=0.009), Gleason score (β=0.358, p < 0.001), and cT_{MRI} (β=0.259, p < 0.001) were significant factors for ECE. For SVI, iPSA (β=0.037, p < 0.001), PPC (β=0.024, p < 0.001), Gleason score (β=0.753, p < 0.001), and cT_{MRI} (β=0.507, p < 0.001) showed statistical significance. BN models to predict ECE and SVI were also successfully established. The overall area under the receiver operating characteristic curve (AUC)/accuracy of the BN models were 0.76/73.0% and 0.88/89.6% for ECE and SVI, respectively. According to internal comparison between the BN model and Roach formula, BN model had improved AUC values for predicting ECE (0.76 vs. 0.74, p=0.060) and SVI (0.88 vs. 0.84, p < 0.001).
Two models to predict pathologic ECE and SVI integrating cT_{MRI} were established and installed on a separate website for public access to guide radiation oncologists.
Adenocarcinoma of the prostate (PCa) is the most common cancer among males from the Unites States, with over 170,000 estimated new cases in 2019 [
In external beam RT for PCa, accurate target volume delineation is crucial due to the routine use of intensitymodulated RT and the increasing utilization of hypofractionated RT, including stereotactic body RT [
The current National Comprehensive Cancer Network (NCCN) risk grouping for PCa is mainly based on clinical staging, initial prostatespecific antigen (iPSA), and Gleason score (GS) obtained from biopsy [
The Bayesian network (BN) model is a statistical framework that represents the conditional dependencies of variables via a directed acyclic graph (DAG). In DAG, nodes indicate clinical variables and edges demonstrate conditional dependencies. The strength of BN is accommodating the heterogeneity between clinical or nonclinical variables and providing interpretable clinical scenarios or probabilities to clinicians. Previous studies have facilitated the use of BN for the prediction of prognosis various in cancers, such as breast cancer and gallbladder cancer [
Thus, in the current study, we sought BN modeling to predict pathologic findings, such as ECE and SVI in patients with PCa. Specifically, in order to guide radiation oncologists for further individualized target contouring based on MRI findings, we established BN models to accurately predict the risk of pathological ECE and SVI by integrating the MRI findings into clinical Tstaging. Furthermore, the positivity rate of the resection margin (RM+ve) was also predicted since these patients are future candidates for salvage RT with increased risk of biochemical relapse [
Patients with PCa confirmed by transrectal ultrasoundguided biopsy and who are undergoing preoperative MRI with radical prostatectomy between 2006 and 2016 were included in the study. All patients were required to have at least eight or more biopsy cores obtained at diagnosis. To minimize patient heterogeneity, those who met the following criteria were excluded: (1) incidentally diagnosed PCa by transurethral prostatectomy for benign prostatic hyperplasia, (2) clinically overt T4 or N1 disease on preoperative imaging, (3) nonadenocarcinoma PCa, (4) history of any antiPCa treatment such as neoadjuvant androgen deprivation therapy prior to radical prostatectomy, and (5) no information regarding iPSA or MRI within 6 months preoperately. A total of 1915 patients met the inclusion and exclusion criteria. Patient characteristics are shown in
All patients were preoperatively evaluated with at least T2weighted and T1weighted images. Diffusionweighted and dynamic contrastenhanced imaging were optional. MRIs were reviewed by 1 of 3 genitourinary imagingspecialized radiologists (S.H.K., J.Y.C., or S.Y.K.). Definite low signal abnormality on T2weighted imaging was regarded as a clinically overt disease (
Multivariate logistic regression analysis (backward stepwise) was performed using SPSS ver. 22.0 (IBM Corp., Armonk, NY). The iPSA, percentage of positive biopsy cores (PPC), both lobe involvement on biopsy (BLIBx), GS, and cT_{MRI} stage were considered as risk factors for predicting ECE, SVI, and RM+ve. All factors were regarded as continuous variables, except BLI on biopsy. GS of 7(4+3) was coded as 7.5 to distinguish from the GS of 7(3+4). The level of statistical significance was set at p < 0.05.
Three aforementioned pathologic findings were defined as class variables: ECE, SVI, and RM+ve. For each class variable, three different BN models were generated [
For each BN model, we validated the model performance by 5fold validation. The mean area under the receiver operating characteristic curve (AUC), calibration index, and accuracy were measured. We visualized the impact of each node on the mean value of the class variable while holding the probability distributions of other variables fixed. Additionally, mutual information (MI) was calculated between each node and the class variable to estimate the probabilistic dependence. MI between two random variables X and Y is defined as the difference between the marginal entropy H(X) and its conditional entropy H(XY) [
The risk of pathological ECE and SVI were calculated in all patients based on the Roach formula as the following: ECE risk (%)=1.5×iPSA+(10×(GS3)); SVI risk (%)=iPSA+(10×(GS6)) [
When cT_{MRI} and pathological Tstage were classified in three groups (T1cT2c, organconfined; T3a, ECE; T3b, SVI), the concordance rate between cT_{MRI} and pathological Tstage was 66.0% (1,293/1,915). Pathological Tstage was downstaged in only 4.8% (92/1,915). However, upstaging was frequently observed with 29.2% of patients (560/1,915) demonstrating higher pathological Tstage compared to cT_{MRI}. Among 110 patients with cT_{MRI}3a, 14 (12.7%) and 36 (32.7%) patients were upstaged to pT3b and downstaged to pT2, respectively. In 150 patients with cT_{MRI}3b, 32 (21.3%) and 24 (16.0%) patients were downstage to pT2 and pT3a, respectively.
In the multivariate logistic regression model, an increase in iPSA, PPC, GS, and cT_{MRI} were significant risk factors for ECE, SVI, and RM+ve. BLIBx was only significant for ECE.
The BN structure predicting the probability of ECE is demonstrated in
The BN model to predict the probability of SVI included cT_{MRI}, GS, iPSA, and PPC (
To predict the probability of RM+ve, we established a BN network model that contained iPSA, PPC, cT_{MRI}, and GS of biopsy (
We implemented each BN model in the webuser interface for convenient use by physicians to estimate the risk of pathological local findings. The web addresses of each model can be found at
In terms of ECE prediction, BN model showed improved performance with marginal significance compared to the Roach formula (AUC, 0.76 vs. 0.74; DeLong’s method p=0.060) (
External beam RT with or without brachytherapy or androgen deprivation therapy is a standard treatment option for PCa throughout all risk groups [
The current international RTtarget delineation guidelines for PCa suggest that ECE and SVI are based on a somewhat simplified risk grouping. The recently published EORTC guidelines suggest adding a 3mm margin around the prostate for the clinical target volume in order to sufficiently cover possible ECE in intermediate/highrisk patients [
Partin’s nomogram as well as Roach’s formula predict the risk of ECE or SVI with respect to possible individual risk factors, such as clinical Tstage, iPSA, and GS [
A major limitation of using the cT_{MRI} concept is that no standardized criteria for local staging with MRI currently exist, which is a major limitation of our study as well. Although we used a simple criterion of definite low signal intensity on T2weighted imaging as an abnormal finding, multiparametric MRI (mpMRI) including diffusionweighted and dynamic contrastenhanced imaging is the recommended modality currently [
BN is a probabilistic graphical model for decision support. In the current study, we could infer important associations between variables and class variables such as ECE, SVI, and RM+ve by using BN models. These BN models can practically support risk adaptedtreatment of patients because they can instantly respond to the clinician’s request or questions with a visual and interpretable representation. In addition, the BN approach can decompose the prediction model and demonstrate how clinical factors affect the predicted class variables. The current study revealed nonlinear as well as linear relationships between risk factors, which are not often captured by the logistic regression model. Moreover, MI was calculated to measure the relative importance among variables. The performance of the BN models in the current study were acceptable for predicting ECE and SVI (
In summary, we have successfully developed two models based on the multivariate logistic regression analysis and BN to estimate the risk of pathologic ECE, SVI, and RM+ve based on several clinical factors in PCa. These models are expected to be applied in patients with PCa to predict the risk of pathological ECE, SVI, and RM+ve, and eventually guide radiation oncologists throughout the targetdelineating process and urologists in deciding treatment modality.
Supplementary materials are available at Cancer Research and Treatment website (
This retrospective study was approved by the Institutional Review Board of the Seoul National University Hospital (IRB No. H1611016805). Informed consent was waived due to the retrospective nature of the study.
Conceived and designed the analysis: Wee CW, Kim JH.
Collected the data: Wee CW, Jeong CW, Kwak C, Kim HH, Ku JH, Kim SH, Cho JY, Kim SY.
Contributed data or analysis tools: Wee CW, Jang BS, Jeong CW, Kwak C, Kim HH, Ku JH, Kim SH, Kim SY.
Performed the analysis: Wee CW, Jang BS, Kim JH.
Wrote the paper: Wee CW, Jang BS, Kim JH.
Conflict of interest relevant to this article was not reported.
A preoperative multiparametric 3.0T magnetic resonance imaging using T1weighted (A), T2weighted (B), diffusionweighted (C), and dynamic contrastenhanced (D) images in a 67yearold male diagnosed as prostate cancer by biopsy. Lesion with suspected seminal vesicle invasion (cT_{MRI}3b) shows low signal intensity on T2 (B) and diffusionweighted (C) images with contrast enhancement (D).
(A) BN structure to estimate the probability of ECE. Each node demonstrates the associated variable, the discretized state, baseline prevalence, mean, deviation values in the study population. (B) The graph showing the impact of each variable on the probability of ECE. xaxis represents the normalized meanvalue of each variable, and yaxis shows the mean probability of ECE. BN, Bayesian network; ECE, extracapsular extension; iPSA, initial prostatespecific antigen; MRI, magnetic resonance imaging.
(A) BN structure to estimate the probability of SVI. Each node demonstrates the associated variable, the discretized state, baseline prevalence, mean, deviation values in study population. (B) The graph showing the impact of each variable on the probability of SVI. xaxis represents the normalized meanvalue of each variable, and yaxis shows the mean probability of SVI. BN, Bayesian network; iPSA, initial prostatespecific antigen; MRI, magnetic resonance imaging; SVI, seminal vesicle invasion.
(A) BN structure to estimate the probability of RM+ve. Each node demonstrates the associated variable, the discretized state, baseline prevalence, mean, deviation values in study population. (B) The graph showing the impact of each variable on the probability of RM+ve. xaxis represents the normalized meanvalue of each variable, and yaxis shows the mean probability of RM+ve. BN, Bayesian network; iPSA, initial prostatespecific antigen; MRI, magnetic resonance imaging; RM+ve, positive resection margin.
Comparison between the predictive accuracy of BN model and Roach formula for pathological ECE (A) and SVI (B) according to the DeLong’s comparison method. AUC, area under the curve; BN, Bayesian network; ECE, extracapsular extension; SVI, seminal vesicle invasion.
Patient and tumor characteristics
Variable  No. (%) 

1,915 (100)  
67 (41–86)  
Below 4  190 (9.9) 
4–10  1,126 (58.8) 
10–20  387 (20.2) 
20 or higher  212 (11.1) 
PPC, median (range, %)  25.0 (5.0–100.0) 
Yes  797 (41.6) 
No  1,076 (56.2) 
Unknown  42 (2.2) 
6  863 (45.1) 
7(3+4)  389 (20.3) 
7(4+3)  297 (15.5) 
8  277 (14.5) 
9  78 (4.1) 
10  11 (0.6) 
1c  243 (12.7) 
2a–2b  718 (37.5) 
2c  694 (36.2) 
3a  110 (5.7) 
3b  150 (7.8) 
ECE  702 (36.7) 
SVI  227 (11.9) 
RM+ve  719 (37.5) 
cT_{MRI}, magnetic resonance imaging–based Tstaging; ECE, extracapsular extension, iPSA, initial prostatespecific antigen; PPC, percentage of positive biopsy cores; RM+ve, positive resection margin; SVI, seminal vesicle invasion; TRUSBx, transrectal ultrasoundguided biopsy.
Clinical stage according to magnetic resonance imaging findings,
Radical prostatectomy specimen.
Results of multivariate logistic regression analysis
Variable  ECE  SVI  RM+ve  



 
β  pvalue  β  pvalue  β  pvalue  
iPSA (ng/mL)  0.050  < 0.001  0.037  < 0.001  0.038  < 0.001 
 
PPC (%)  0.033  < 0.001  0.024  < 0.001  0.021  < 0.001 
 
BLIBx  0.359  0.009  −0.003  0.988  −0.177  0.161 
 
Gleason scores on biopsy  0.358  < 0.001  0.753  < 0.001  0.131  0.037 
 
cT_{MRI} stage 
0.259  < 0.001  0.507  < 0.001  0.149  0.006 
 
Constant  −5.486  < 0.001  −10.459  < 0.001  −2.894  < 0.001 
cT_{MRI}, magnetic resonance imaging–based Tstaging; BLIBx, both lobe involvement on biopsy; ECE, extracapsular extension; iPSA, initial prostatespecific antigen; PPC, percentage of positive biopsy cores; RM+ve, positive resection margin; SVI, seminal vesicle invasion.
Clinical stage according to magnetic resonance imaging findings.
Performance of BN models
BN models  5Fold validation metrics  

Overall AUC 
Overall Calibration Index (%)  Overall accuracy (%)  
ECE  0.76  79.7  73.0 
SVI  0.88  62.6  89.6 
RM+ve  0.70  75.1  69.1 
AUC, area under the curve; BN, Bayesian network; ECE, extracapsular extension, RM+ve, positive resection margin; SVI, seminal vesicle invasion.
Interpretation: 0.5, no discrimination; 0.7–0.8, acceptable; 0.8–0.9, excellent; > 0.9, outstanding [
Node significance to the class variable
Node  Normalized mutual information (%)  Relative mutual information (%)  Relative significance  pvalue 

PPC  8.22  8.67  1.000  < 0.001 
iPSA (ng/mL)  7.74  8.17  0.942  < 0.001 
Gleason scores of biopsy  7.56  7.97  0.920  < 0.001 
cT_{MRI} stage 
6.63  6.99  0.807  < 0.001 
Both lobe involvement  2.62  2.76  0.319  < 0.001 
cT_{MRI} stage 
9.85  18.75  1.000  < 0.001 
Gleason scores of biopsy  9.02  17.17  0.916  < 0.001 
iPSA (ng/mL)  7.25  13.81  0.736  < 0.001 
PPC  7.15  13.61  0.726  < 0.001 
iPSA (ng/mL)  6.40  6.70  1.000  < 0.001 
PPC  5.26  5.51  0.822  < 0.001 
cT_{MRI} stage 
3.76  3.94  0.588  < 0.001 
Gleason scores of biopsy  3.57  3.74  0.559  < 0.001 
cT_{MRI}, magnetic resonance imaging based Tstaging; ECE, extracapsular extension; iPSA, initial prostatespecific antigen; MRI, magnetic resonance imaging; PPC, percentage of positive biopsy cores; RM+ve, positive resection margin; SVI, seminal vesicle invasion.
Gtest,
Clinical stage according to magnetic resonance imaging findings.