Purpose: To investigate whether prebiopsy multi-parametric (mp) MRI can help to improve predictive performance in prostate cancer.

Experimental Design: Based on a support vector machine (SVM) analysis, we prospectively modeled clinical data (age, PSA, digital rectal examination, transrectal ultrasound, PSA density, and prostate volume) and mp-MRI findings [Prostate Imaging and Reporting and Data System (PI-RADS) score and tumor–node–metastasis stage] in 985 men to predict the risk of prostate cancer. The new nomogram was validated in 493 patients treated at the same institution. Multivariable Cox regression analyses assessed the association between input variables and risk of prostate cancer, and area under the receiver operating characteristic curve (Az) analyzed the predictive ability.

Results: At 5-year follow-up period, 34.3% of patients had systemic progression of prostate cancer. Nomogram (SVM-MRI) predicting 5-year prostate cancer rate trained with clinical and mp-MRI data was accurate and discriminating with an externally validated Az of 0.938, positive predictive value (PPV) of 77.4%, and negative predictive value of 91.5%. The improvement was significant (P < 0.001) compared with the nomogram trained with clinical data. When stratified by PSA, SVM-MRI nomogram had high PPV (93.6%) in patients with PSA > 20 ng/mL, with intermediate to low PPV in PSA 10 to 20 ng/mL (64%), PSA 4 to 10 ng/mL (55.8%), and PSA 0 to 4 ng/mL (29%). PI-RADS score (Cox HR, 2.112; P < 0.001), PSA level (HR, 1.435; P < 0.001), and age (HR, 1.012; P = 0.043) were independent predictors of prostate cancer.

Conclusions: Featured with low false positive rate, mp-MRI could be the first investigation of a man with a raised PSA before prostate biopsy. Clin Cancer Res; 23(14); 3692–9. ©2017 AACR.

Translational Relevance

This prospective study, with the largest patient capacity in China, sought to determine whether it is time to consider a role for MRI before prostate biopsy. With features such as high accuracy and low false positive rate, the newly developed nomogram based on prebiopsy clinical and imaging data can help to predict the clinical outcome accurately in 92.8% patients, even without biopsy results. This finding strongly supports the theory that prostate multi-parametric MRI could be the first investigation of a man with a raised PSA before invasive biopsy, avoiding overdiagnosis and overtreatment.

Prostate cancer is the most commonly diagnosed cancer that affects elderly men and therefore is a bigger health concern in developed countries (1), and now the incidence is also rapidly increasing in China (2). The use of PSA in combination with digital rectal examination (DRE) is a widely adopted, population-based screening program in clinical practice that can effectively detect prostate cancer at earlier, asymptomatic stages (3, 4). However, this traditional screening plan has recently come in for a lot of criticism due to its noticeable limitations for cancer detecting, disease monitoring, and patient management (5). For its natural feature of variability, PSA-based screening leads to an increase in the number of unnecessary biopsies and high risk of overtreatment (6, 7). Now, there is a lack of evidence to support that PSA-based screening can influence prostate cancer mortality. And a Prostate, Lung, Colorectal and Ovarian Cancer Screening program has been discredited because of poor survival advantage (8). Therefore, there is an imminent need for simplified predictive tools that include novel pretreatment variables that can extend the clinical performance of traditional PSA-based nomogram.

Over the past decades, multi-parametric MRI (mp-MRI) was only recommended for prostate cancer staging after biopsy (9, 10). Recently, it has resulted in great ability in determining tumorous aggressiveness and, potentially, predicting the risk of disease progression or recurrence (11–14). Many urologists now request a pelvic MRI scan as the first investigation of a man with a raised PSA before the prostate biopsy (15, 16). This allows men with a normal scan to be monitored, and if the men with an abnormal scan, this sensible plan can permit better-informed decisions to be made about recommendations for prostate biopsy, watchful waiting, radiotherapy, or prostatectomy (3, 9, 17). These findings suggested that mp-MRI could do more in entire clinical workflow for patient management.

Therefore, the purpose of this study is to investigate whether prebiopsy mp-MRI findings, when converted into a prognostic nomogram, can help to improve predictive performance in a large contemporary cohort of Chinese patients undergoing PSA screening, with a focus on 5-year prostate cancer rate.

Patient population and follow-up

This study was approved by our local Institutional Review Board (Peking University First Hospital, Beijing, China). Between July 2002 and December 2009, 1,778 consecutive patients with elevated PSA undergoing a prebiopsy prostate mp-MRI were prospectively recruited. Written informed consents were obtained from all patients. Before MR examination, a questionnaire survey, including the latest PSA level within 1 month, DRE reports, transrectal ultrasound (TRUS) findings, previous biopsies, previous prostate MRI scans, and history of previous prostate treatment or intervention, was conducted.

Patients were consistently followed with intervals of 6 months to 2 years and were censored at the occurrence of prostate cancer, emigration, or December 31, 2014, whichever came first. The follow-up included measurements of serum PSA level, and/or a DRE, and/or MR examination, and/or an invasive TRUS-based biopsy. At the end of follow-up period, prostate cancer was determined by histopathologic examination, i.e., TRUS-based biopsy, transurethral resection of the prostate (TURP), and radical prostatectomy (RP). Patients with initial negative biopsy were specially followed by PSA, DRE, MRI, and/or repeated TRUS biopsy until the end of follow-up period. Patients without histopathologic results were treated as prostate cancer if they had consistently elevated PSA level (>10 ng/mL after three rounds of PSA testing) and/or significantly positive findings on imaging examination (i.e., CT, MR, or bone scan). The exclusion criteria included: (i) patient underwent prostate biopsy before mp-MRI (n = 6, 0.33%); (ii) patients failed to receive standardized MR or underwent the MR examination from outside institutions (n = 3, 0.17%); (iii) patients with inadequate information that cannot determine the clinical outcome within 5 years following MR examination (n = 291, 16.4%). At last, 1,478 patients (median age, 68 years; age range, 56–81 years) were eligible for clinical evaluation.

MR examination

All imaging examinations were performed with 1.5-T MR scanners (Signa HDxt, GE Medical System) using pelvic phased-array coils. As per the standard prostatic MRI at our institution, the protocol included a combination of T2-weighted imaging, diffusion-weighted imaging, MR spectroscopy imaging, and dynamic contrast-enhanced MR imaging. The MR parameters and representative MR imaging cases are presented in Supplementary Data S1 and S2.

Imaging analysis

The prostate MR images were retrospectively interpreted by two radiologists with 3 years (R. Wang) and 10 years (X. Wang) of experience who were blinded to the histologic results and all clinical information. In each patient, the radiologists first identified the most suspicious cancer lesion (leading lesion) in the peripheral zone (PZ) and/or the transition zone (TZ). For the patient with multiple lesions, the largest sized lesion, and/or dominantly low signal intensity on apparent diffusion coefficient (ADC) maps, and/or suspected extracapsular extension (ECE), and/or suspected seminal vesicle invasion (SVI) was defined as the leading lesion (Supplementary Data S3). The following MR features were then recorded according to the guidelines of European Society of Urogenital Radiology (9): (i) tumor location (PZ or TZ); (ii) a Prostate Imaging and Reporting and Data System (PI-RADS) score using v2 (18); (iii) MR-detected ECE, SVI, local lymph node (LN), and local bone metastasis. During the image interpretation, any inter-reader disagreement in qualitative assessments was discussed until consensus was reached. This procedure was performed after a statistical measurement of inter-observer variability.

Model development

The patients were divided into two groups in a randomized fashion: one group was designated the training group (n = 985) for training the predictive model, and the other group was designated the validation group (n = 493) for the evaluation of the accuracy of the predictive model. This randomized fashion for the classification of patients is an advisable way to avoid model overfitting. Patient age (<60 years and ≥60 years of age), DRE findings (positive or negative), TRUS results (positive or negative), prostate volume, PSA level, PSA density (PSAD = PSA/prostate volume), and MRI findings, including PI-RADS score, ECE, SVI, local LN, and local bone metastasis, were included as input data. The prostate volume was determined by TRUS or MRI, wherein the volume is the products of height (H), width (W), and length (L) of prostate gland, as well as with a coefficient of 0.52 (19).

Predictive model was constructed from data of the training group composed of 985 patients by using a novel support vector machine (SVM) analysis. SVM is a supervised machine learning technique used for classification and regression analysis. SVM algorithm tries to construct an optimal separating hyperplane that maximizes the margin, where the margin is the largest distance to the nearest training data point of any class (20, 21). Unlike the traditional artificial neural networks, SVM does not have to undertake a trial and error parameter decision process, while determines optimal performance conditions automatically if the kernel type is set. In this study, SVM with radial basis function (RBF) kernel was applied to resolve the two class problems, clinical progression of prostate cancer or not (non–prostate cancer). A RBF kernel, K, maps the original data with the kernel function as:

formula

where x and t are two feature vectors, and Gamma (g) controls the shape of the decision hyperplane. As the direct output value of SVM analysis does not show probabilities of prostate cancer, we converted their output values to the probabilities (Pi) by applying a sigmoid function as follows:

formula

where x is the output value of SVM analysis. The value of Pi indicates the probabilities that the patient has prostate cancer.

In the first model (SVM-clinical), six input parameters, i.e., patient age, PSA level, prostate volume, PSAD, DRE, and TRUS findings, were used. In the second model (SVM-MRI), the mp-MRI findings (PI-RADS score, ECE, SVI, local LN, and bone metastasis) were added.

Statistical analysis

Regarding the reproducibility of MR measurements, the qualitative results (i.e., PI-RADS score, ECE, or SVI) between the two readers were assessed by kappa (κ) statistics test.

The performance of two predictive models was analyzed from 493 validation data, by using a binary ROC regression analysis and quantified by the areas under the ROC curves (Az). The predictive sensitivity (SEN), specificity (SPE), positive predictive value (PPV), negative predictive value (NPV), and overall accuracy (ACC) were calculated at a cutoff point that maximized the value of the Youden index. In order to determine the independent risk factors who were associated with the time to progression of prostate cancer, a Kaplan–Meier analysis with the log-rank test was firstly employed to identify the univariate associations between the predictors and progression of prostate cancer. Factors significant in the Kaplan–Meier test were further analyzed by using a multivariable Cox regression analysis to determine the independent risk factors who were associated with the time to progression of prostate cancer. Statistical analysis was performed with commercially available software (SPSS 15.0 for Windows; SPSS). A P value of <0.05 was considered statistically significant.

Table 1 lists clinical and MRI characteristics of the training and validation cohorts. As of December 2014, overall 507 (34.3%) of the 1,478 patients in the study had a 5-year clinical progression of prostate cancer. The TRUS biopsy was performed initially in 429 patients and produced a 79.0% (339/429) positive detection rate. The remaining 168 patients with prostate cancer were detected by repeated biopsy (n = 62), comprehensive clinical-imaging analysis (n = 101), and TURP incidentally (n = 5). The cumulative rate of prostate cancer was 29.6% (437/1,478) at 1 year, 31.8% (470) at 3 years, and 34.3% (507) at 5 years, respectively. Among of all positive cases, 47.1% (239/507) patients underwent endocrine therapy, 13.2% (67/507) underwent prostate androgen deprivation therapy, 18.3% (93/507) underwent active RP, 15.2% (77/507) underwent radiotherapy, and 2.2% (11/507) underwent chemotherapy. During the 5-year follow-up period, 7.1% (36/507) patients died of prostate cancer and 12.8% (65/507) died of another cause.

Table 1.

Clinical and MRI characteristics of patients in training and validation cohorts

VariableTraining data (n = 985)Validation data (n = 493)P
Clinical characteristics 
 Median age (IQR) 70 (65–75) 70 (65–75) 0.967 
 Median follow-up (IQR) 40 (1–80) 38 (1–78) 0.936 
 Median (ng/mL) PSA (IQR) 10.3 (6.2–19.1) 9.8 (6.2–18.5) 0.981 
 Median prostate volume (IQR) 54.5 (35.6–73.1) 51.9 (32.5–63.7) 0.463 
 Median PSAD 0.19 (0.10–0.40) 0.20 (0.11–0.44) 0.897 
 DRE positive, n (%) 82 (8.3%) 58 (11.8%) 0.033 
 TRUS positive, n (%) 73 (7.4%) 30 (6.1%) 0.387 
 Initial biopsy positive, n (%) 232 (42.6%) 107 (48.0%) 0.175 
Diagnosis of prostate cancer 
 Total detecting rate, n (%) 343 (34.8%) 164 (33.3%) 0.255 
 Detected by histopathology, n (%) 294 (29.8%) 107 (21.7%)  
 Diagnosed by clinical analysis, n (%) 49 (5.0%) 57 (11.6)  
MR findings 
 Median PI-RADS score (IQR) 2 (1–4) 2 (1–4) 0.721 
 ECE positive, n (%) 188 (19.1%) 98 (19.9%) 0.727 
 SVI positive, n (%) 117 (11.9%) 20 (4.1%) 0.001 
 Local LN invasion, n (%) 49 (5.0%) 24 (4.9%) 0.929 
 Local bone metastasis, n (%) 84 (8.5%) 29 (5.9%) 0.073 
Final Gleason score 
 ≤3+3, n (%) 42 (14.3%) 10 (9.3%) 0.415 
 3+4, n (%) 56 (19.0%) 23 (21.5%) 0.876 
 4+3, n (%) 26 (8.8%) 10 (9.3%) 0.924 
 ≥4+4, n (%) 170 (57.8%) 64 (59.8%) 0.931 
Treatmenta and death rate 
 ET, n (%) 183 (18.6%) 56 (11.3%)  
 ADT, n (%) 54 (5.5%) 13 (2.6%)  
 RP, n (%) 65 (6.6%) 27 (5.5%)  
 RT, n (%) 50 (5.1%) 27 (5.5%)  
 Chemotherapy, n (%) 5 (0.5%) 6 (1.2%)  
 Total death rate, n (%) 73 (7.4%) 28 (5.6%) 0.087 
 Death by prostate cancer, n (%) 25 (2.5%) 11 (2.2%) 0.488 
VariableTraining data (n = 985)Validation data (n = 493)P
Clinical characteristics 
 Median age (IQR) 70 (65–75) 70 (65–75) 0.967 
 Median follow-up (IQR) 40 (1–80) 38 (1–78) 0.936 
 Median (ng/mL) PSA (IQR) 10.3 (6.2–19.1) 9.8 (6.2–18.5) 0.981 
 Median prostate volume (IQR) 54.5 (35.6–73.1) 51.9 (32.5–63.7) 0.463 
 Median PSAD 0.19 (0.10–0.40) 0.20 (0.11–0.44) 0.897 
 DRE positive, n (%) 82 (8.3%) 58 (11.8%) 0.033 
 TRUS positive, n (%) 73 (7.4%) 30 (6.1%) 0.387 
 Initial biopsy positive, n (%) 232 (42.6%) 107 (48.0%) 0.175 
Diagnosis of prostate cancer 
 Total detecting rate, n (%) 343 (34.8%) 164 (33.3%) 0.255 
 Detected by histopathology, n (%) 294 (29.8%) 107 (21.7%)  
 Diagnosed by clinical analysis, n (%) 49 (5.0%) 57 (11.6)  
MR findings 
 Median PI-RADS score (IQR) 2 (1–4) 2 (1–4) 0.721 
 ECE positive, n (%) 188 (19.1%) 98 (19.9%) 0.727 
 SVI positive, n (%) 117 (11.9%) 20 (4.1%) 0.001 
 Local LN invasion, n (%) 49 (5.0%) 24 (4.9%) 0.929 
 Local bone metastasis, n (%) 84 (8.5%) 29 (5.9%) 0.073 
Final Gleason score 
 ≤3+3, n (%) 42 (14.3%) 10 (9.3%) 0.415 
 3+4, n (%) 56 (19.0%) 23 (21.5%) 0.876 
 4+3, n (%) 26 (8.8%) 10 (9.3%) 0.924 
 ≥4+4, n (%) 170 (57.8%) 64 (59.8%) 0.931 
Treatmenta and death rate 
 ET, n (%) 183 (18.6%) 56 (11.3%)  
 ADT, n (%) 54 (5.5%) 13 (2.6%)  
 RP, n (%) 65 (6.6%) 27 (5.5%)  
 RT, n (%) 50 (5.1%) 27 (5.5%)  
 Chemotherapy, n (%) 5 (0.5%) 6 (1.2%)  
 Total death rate, n (%) 73 (7.4%) 28 (5.6%) 0.087 
 Death by prostate cancer, n (%) 25 (2.5%) 11 (2.2%) 0.488 

Abbreviations: ADT, androgen deprivation therapy; ET, endocrine therapy; IQR, interquartile range; RT, radiation therapy.

an is the therapy frequency (1 patient may receive more than one therapy).

Supplementary Data S4 show the workflow of machine learning analysis with SVM for model development in 985 training data. Figure 1 shows the performance of SVM-clinical and SVM-MRI model in training data and validation data, respectively. Using traditional clinical input parameters, i.e., patient age, PSA level, prostate volume, PSAD, DRE, and TRUS findings, the Az of SVM-clinical model for prostate cancer prediction was 0.716 [95% confidence interval (CI), 0.686–0.744] for training data and 0.715 (95% CI, 0.673–0.754) for validation data. With addition of mp-MRI findings, the Az (training: 0.930, 95% CI, 0.912–0.945; validation: 0.932, 95% CI, 0.906–0.953) of SVM-MRI model becomes significantly larger (P < 0.001).

Figure 1.

The performance of two predictive nomograms, i.e., SVM-clinical versus SVM-MRI, in prediction of prostate cancer in training data (A) and validation data (B).

Figure 1.

The performance of two predictive nomograms, i.e., SVM-clinical versus SVM-MRI, in prediction of prostate cancer in training data (A) and validation data (B).

Close modal

Table 2 shows the internal validity of two predictive models, i.e., SVM-clinical versus SVM-MRI, to predict 5-year prostate cancer rate for different cutoff points in the 985 modeling and 493 validation data, respectively. Figure 2 shows 1-, 3-, and 5-year prediction rate functions (PRF) of prostate cancer estimated with SVM-clinical and SVM-MRI model in 493 validation data. The cumulative predictive accuracy estimated by SVM-MRI model was 92.8% at 1 year, 87.9% at 3 years, and 84.6% at 5 years, respectively, significantly higher than those estimated by SVM-clinical model (68.9%, 67.9%, and 68.6%, respectively). It shows that the PRF constructed by SVM-MRI model is excellently close to patients' actuarial PRF until 32 months after the follow-up. While both two models have increased bias with long-term follow-up period. When stratified by PSA level, the SVM-MRI nomogram had high PPV (93.6%) in patients with PSA > 20 ng/mL, while with intermediate to low PPVs in PSA 10 to 20 ng/mL (64%), PSA 4 to 10 ng/mL (55.8%), and PSA 0 to 4 ng/mL (29%). The SVM-MRI model had an NPV of higher than 90% in all four groups with different PSA level (Fig. 3).

Table 2.

Internal validity of the SVM-clinical/SVM-MRI model to predict 5-year prostate cancer rate for different cutoff points in the validation data set (n = 493)

Performance parameter for SVM-clinical/SVM-MRI model
Cutoff levelSEN (%)SPE (%)PPV (%)NPV (%)ACC (%)
>0.1 97.0/95.7 6.1/50.8 33.9/49.2 80.0/95.9 36.3/65.7 
>0.3 84.8/93.3 37.9/78.7 40.5/68.6 83.3/95.9 53.5/83.6 
>0.5a 27.4/83.5 89.9/87.8 57.7/77.4 71.3/91.5 69.2/86.4 
>0.7 4.3/56.7 99.4/99.4 77.8/97.9 67.6/82.2 67.7/85.2 
Performance parameter for SVM-clinical/SVM-MRI model
Cutoff levelSEN (%)SPE (%)PPV (%)NPV (%)ACC (%)
>0.1 97.0/95.7 6.1/50.8 33.9/49.2 80.0/95.9 36.3/65.7 
>0.3 84.8/93.3 37.9/78.7 40.5/68.6 83.3/95.9 53.5/83.6 
>0.5a 27.4/83.5 89.9/87.8 57.7/77.4 71.3/91.5 69.2/86.4 
>0.7 4.3/56.7 99.4/99.4 77.8/97.9 67.6/82.2 67.7/85.2 

NOTE: The cutoff level is the level of predictive probability (Pi). There are the points (the cutoff levels) over the receiver operator curve. For example, accuracy of 80% for a cutoff level of 0.5 means that the accuracy of the predictive model with a 50% probability or more for prediction of prostate cancer is 80%.

aUsing the cutoff value of 0.5, SVM-MRI has significantly (P < 0.001) higher SEN, PPV, NPV, and ACC than SVM-clinical model.

Figure 2.

One-year (A), 3-year (B), and 5-year (C) PRFs of prostate cancer by SVM-clinical and SVM-MRI model in 493 validation data. It shows that SVM-MRI is dominantly superior to SVM-clinical model in determination of patients' short-term outcome (less than 32 months, red arrow).

Figure 2.

One-year (A), 3-year (B), and 5-year (C) PRFs of prostate cancer by SVM-clinical and SVM-MRI model in 493 validation data. It shows that SVM-MRI is dominantly superior to SVM-clinical model in determination of patients' short-term outcome (less than 32 months, red arrow).

Close modal
Figure 3.

The predictive performance of SVM-MRI model in four groups stratified by PSA level.

Figure 3.

The predictive performance of SVM-MRI model in four groups stratified by PSA level.

Close modal

Unadjusted and multivariable adjusted HRs with cox proportional hazards model are presented in Supplementary Data S5. In univariate analyses, age, baseline PSA level, prostate volume, and PI-RADS score were all significantly associated with an increased risk of prostate cancer. In adjusted analysis, PI-RADS score had the highest Cox HR (2.112; P < 0.001) for prediction of time to prostate cancer, followed by PSA level (HR, 1.435; P < 0.001) and age (HR, 1.012; P = 0.043). The distribution of PI-RADS scores predicting the probability of prostate cancer is shown in Table 3, showing the incidence of prostate cancer increased with the mp-MRI score. The vast majority of patients (512/1,478, 34.6%) were given a score of 4 and 5, signifying an mp-MRI report of an organ-confined tumor. The incidence of prostate cancer was 2.9% in whom no tumor was apparent (score, 1–2) and 3.5% in whom prostate cancer could not be excluded (scores 3). Using a cutoff value of score 3, PI-RADS produced good predictive ability in prostate cancer, with a SEN of 81.3%, SPE of 89.7%, PPV of 80.5%, and NPV of 90.2%, respectively. Unilateral or bilateral ECE was reported in 286 of 1,478 (19.4%) patients, SVI was reported in 137 of 1,478 (9.3%) patients, local LN metastasis was reported in 73 of 1,478 (4.9%) patients, and local bone metastasis was reported in 113 of 1,478 (7.6%) patients.

Table 3.

Distribution of the 1,478 patients according to PI-RADS score, with actual recorded events of prostate cancer for each score

PI-RADSDistribution, N (%)Actual prostate cancer, N (%)SEN, %SPE, %PPV, %NPV, %Cutoff
Score 1 576 (39.0%) 30 (2.0%) 94.1 56.2 52.9 94.8 >1 
Score 2 240 (16.2%) 13 (0.9%) 91.5 79.6 70.1 94.7 >2 
Score 3 150 (10.1%) 52 (3.5%) 81.3 89.7 80.5 90.2 >3 
Score 4 222 (15.0%) 129 (8.7%) 55.8 99.3 97.6 81.1 >4 
Score 5 290 (19.6%) 283 (19.1%)      
PI-RADSDistribution, N (%)Actual prostate cancer, N (%)SEN, %SPE, %PPV, %NPV, %Cutoff
Score 1 576 (39.0%) 30 (2.0%) 94.1 56.2 52.9 94.8 >1 
Score 2 240 (16.2%) 13 (0.9%) 91.5 79.6 70.1 94.7 >2 
Score 3 150 (10.1%) 52 (3.5%) 81.3 89.7 80.5 90.2 >3 
Score 4 222 (15.0%) 129 (8.7%) 55.8 99.3 97.6 81.1 >4 
Score 5 290 (19.6%) 283 (19.1%)      

To our knowledge, this is the first prospective study to investigate the predictive role of prebiopsy mp-MRI for prostate cancer with a long-term follow-up in 1,478 consecutive Chinese patients. We report that, at 5-year follow-up period, 34.3% of patients had systemic progression of prostate cancer, and 29.6% was detected within initial 1-year follow-up, whereas only 2.5% was detected within the following 4-year period. In addition, we noted that the machine learning–based model using the input variables of clinical variables and mp-MRI findings (SVM-MRI) shows high SEN (83.5%) and SPE (87.8%) to predict prostate cancer statistically more accurate (86.4%) than PSA-based screening program. And importantly, high NPVs (larger than 90% in any PSA level) made the mp-MRI extraordinarily useful for initial evaluation before an invasive biopsy.

The conventional diagnostic tools for detection of prostate cancer are DRE, PSA, and TRUS-guided biopsies (3), whereas mp-MRI is commonly recommended for cancer staging after biopsies. However, in about 18% of all patients, prostate cancer is detected by a suspect DRE alone, irrespective of the PSA level (22). A suspect DRE in patients with a PSA level of up to 2 ng/mL has a PPV of 5% to 30% (23, 24). In our study, only 9.5% of all patients was detected positively by DRE alone, and 7.0% by TRUS alone. A suspect DRE or TRUS in patients with a PSA level of up to 4 ng/mL has a PPV of 36.7%. To address these limitations, we proposed a prognostic nomogram to more accurately quantify the individual probability of prostate cancer based on a machine learning approach. We noted that, even loading an advanced SVM analysis, the model constructed from clinical input variables only increased the PPV to 40.5%, and overall performance (Az) was 0.715. When mp-MRI was added, the Az increased to 0.938, and overall predictive accuracy increased to 86.4%. In addition, the new nomogram increased NPVs from 71.3% to 91.5%. By this improvement, 20.2% patients with PSA testing positive could avoid undergoing invasive biopsy. This result confirms strongly the thesis that mp-MR findings can complement the standard clinical nomogram for the improvement of the predictive performance. The nomogram by integrating mp-MRI and clinical variables could be noninvasive and prospective, and thus allows for immediate identification of the patients accounted for biopsy or not. This thesis now receives more support in the urologic literatures (25–27).

The significant contribution of mp-MRI to predict prostate cancer may be explained by the assumption that PI-RADS score serves as a possible indicator for the probability of prostate cancer (28–30). We noted that PI-RADS score > 3 resulted in only 9.8% false negative results. And among all mp-MRI parameters, PI-RADS score was the only independent predictor of detection rate of prostate cancer, as well as with the highest Cox HR value among all independent predictors. However, mp-MRI has relatively low ability in detection of clinically insignificant prostate cancer. As most of insignificant tumors have volume less than 0.5 mL, PSA level less than 10 ng/mL, or Gleason score less than 3+3, which are commonly invisible on mp-MR images, we observed that, among 816 patients who had a PI-RADS score 1 and score 2 (MR-invisible), the rate of prostate cancer was 5.3% (43/816). Among patients who had a PSA level less than 4 ng/mL and a PSA level between 4 and 10 ng/mL, the rate of prostate cancer was 6.4% (12/188) and 19.4% (104/535), respectively. The PPV of SVM-MRI model for predicting prostate cancer in this two groups was only 29.0% and 55.8%, respectively. While among 351 patients who had PSA larger than 20 ng/mL, the PPV increased to 93.6%. However, we noted that the SVM-MRI model had high NPVs (>90%) regardless of PSA level. This finding strongly supports that mp-MRI can help to detect the most biologically significant prostate cancer and predict all negative cases, and thus allows to immediately identify the patients who are accounted for active RP treatment and who only need an active surveillance.

We noted that among all studied patients, the first 1-year true positive rate is 29.6% (n = 437), and the first 1-year cumulative predictive accuracy by SVM-MRI model is as high as 92.8%. This indicates the outcome of most patients who underwent PSA screening program could be accurately predicted in the first 1-year follow-up period if they received a prebiopsy mp-MRI examination, even without invasive TRUS biopsy. And second, we noted the true positive rate is only 2.5% (n = 70) in the following 4-year follow-up period. This indicates that there is limited clinical significance for a long-term follow-up in patients who are detected negatively in the first 1-year period.

A couple of limitations of our study warrant mention. First, mp-MRI data were obtained at a 1.5 T MR scanner with utilization of a surface coil. Although at present, MRI at 3 T with quantitative parameters, i.e., ADCs, Ktrans, and texture variables, is often considered state of the art. Second, because of so long-term periods, the imaging criteria for prostate cancer have been greatly changed (i.e., from early experience-based norms to a structured PI-RADS v1 and recently updated PI-RADS v2), the decision on which may make influence on patient management. Moreover, the new PI-RADS v2, even showing great advantage, still needs to be validated by numerous multicenter studies in the future. Third, because we did not have a device that supports the TRUS/MR fusion-guided in-bore biopsy, the histopathology validation is not with a targeted approach, which may disqualify the MRI read outs. Last, this is a single center study, and an external validation cohort should be included to test the reproducibility of established model in future.

We first investigated the systemic outcome of a crowd underwent PSA-based screening and prebiopsy mp-MRI, and demonstrated the predictive role of prebiopsy mp-MRI for prostate cancer by using an advanced machine learning–based approach. Here, we answer two important questions at the beginning of the article: (1) Machine learning analysis of mp-MRI findings can help to improve the predictive performance in prostate cancer, by which, the outcome of 92.8% patients could be accurately predicted in the first 1-year follow-up period by combining clinical and mp-MRI findings. (2) Advanced in high PPVs and NPVs, prostate mp-MRI could be the first investigation of a man with a raised PSA before invasive biopsy, avoiding overdiagnosis and overtreatment.

No potential conflicts of interest were disclosed.

Conception and design: R. Wang, J. Wang, Y.-D. Zhang, X. Wang

Development of methodology: R. Wang, J. Wang, J. Hu, Y.-D. Zhang, X. Wang

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): R. Wang, J. Hu, Y. Jiang, Z. Zhao, Y.-D. Zhang, X. Wang

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): R. Wang, J. Wang, G. Gao, Y. Jiang, Z. Zhao, Y.-D. Zhang, X. Wang

Writing, review, and/or revision of the manuscript: R. Wang, J. Wang, Y.-D. Zhang, X. Wang

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): R. Wang, G. Gao, Y. Jiang, Z. Zhao, X. Zhang, Y.-D. Zhang, X. Wang

Study supervision: R. Wang, Y.-D. Zhang, X. Wang

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.

1.
Siegel
RL
,
Miller
KD
,
Jemal
A
. 
Cancer statistics, 2016
.
CA Cancer J Clin
2016
;
66
:
7
30
.
2.
Chen
W
,
Zheng
R
,
Baade
PD
,
Zhang
S
,
Zeng
H
,
Bray
F
, et al
Cancer statistics in China, 2015
.
CA Cancer J Clin
2016
;
66
:
115
32
.
3.
Mottet
N
,
Bellmunt
J
,
Bolla
M
,
Briers
E
,
Cumberbatch
MG
,
De Santis
M
, et al
EAU-ESTRO-SIOG Guidelines on Prostate Cancer. Part 1: Screening, diagnosis, and local treatment with curative intent
.
Eur Urol
2016
.
pii: S0302-2838
30470
5
.
4.
Hayes
JH
,
Barry
MJ
. 
Screening for prostate cancer with the prostate-specific antigen test: A review of current evidence
.
JAMA
2014
;
311
:
1143
9
.
5.
Chou
R
,
Croswell
JM
,
Dana
T
,
Bougatsos
C
,
Blazina
I
,
Fu
R
, et al
Screening for prostate cancer: A review of the evidence for the U.S. Preventive Services Task Force
.
Ann Intern Med
2011
;
155
:
762
71
.
6.
Schroder
FH
,
Hugosson
J
,
Roobol
MJ
,
Tammela
TL
,
Ciatto
S
,
Nelen
V
, et al
Screening and prostate-cancer mortality in a randomized European study
.
N Engl J Med
2009
;
360
:
1320
8
.
7.
Zhu
X
,
Albertsen
PC
,
Andriole
GL
,
Roobol
MJ
,
Schroder
FH
,
Vickers
AJ
. 
Risk-based prostate cancer screening
.
Eur Urol
2012
;
61
:
652
61
.
8.
Andriole
GL
,
Crawford
ED
,
Grubb
RL
 3rd
,
Buys
SS
,
Chia
D
,
Church
TR
, et al
Mortality results from a randomized prostate-cancer screening trial
.
N Engl J Med
2009
;
360
:
1310
9
.
9.
Barentsz
JO
,
Richenberg
J
,
Clements
R
,
Choyke
P
,
Verma
S
,
Villeirs
G
, et al
ESUR prostate MR guidelines 2012
.
Eur Radiol
2012
;
22
:
746
57
.
10.
Heidenreich
A
,
Bastian
PJ
,
Bellmunt
J
,
Bolla
M
,
Joniau
S
,
van der Kwast
T
, et al
EAU guidelines on prostate cancer. part 1: Screening, diagnosis, and local treatment with curative intent-update 2013
.
Eur Urol
2014
;
65
:
124
37
.
11.
Yakar
D
,
Debats
OA
,
Bomers
JG
,
Schouten
MG
,
Vos
PC
,
van Lin
E
, et al
Predictive value of MRI in the localization, staging, volume estimation, assessment of aggressiveness, and guidance of radiotherapy and biopsies in prostate cancer
.
J Magn Reson Imaging
2012
;
35
:
20
31
.
12.
Jung
SI
,
Donati
OF
,
Vargas
HA
,
Goldman
D
,
Hricak
H
,
Akin
O
. 
Transition zone prostate cancer: Incremental value of diffusion-weighted endorectal MR imaging in tumor detection and assessment of aggressiveness
.
Radiology
2013
;
269
:
493
503
.
13.
Vargas
HA
,
Akin
O
,
Franiel
T
,
Mazaheri
Y
,
Zheng
J
,
Moskowitz
C
, et al
Diffusion-weighted endorectal MR imaging at 3 T for prostate cancer: Tumor detection and assessment of aggressiveness
.
Radiology
2011
;
259
:
775
84
.
14.
Schoots
IG
,
Petrides
N
,
Giganti
F
,
Bokhorst
LP
,
Rannikko
A
,
Klotz
L
, et al
Magnetic resonance imaging in active surveillance of prostate cancer: A systematic review
.
Eur Urol
2015
;
67
:
627
36
.
15.
Numao
N
,
Yoshida
S
,
Komai
Y
,
Ishii
C
,
Kagawa
M
,
Kijima
T
, et al
Usefulness of pre-biopsy multiparametric magnetic resonance imaging and clinical variables to reduce initial prostate biopsy in men with suspected clinically localized prostate cancer
.
J Urol
2013
;
190
:
502
8
.
16.
Patel
U
,
Dasgupta
P
,
Challacombe
B
,
Cahill
D
,
Brown
C
,
Patel
R
, et al
Pre-biopsy 3-Tesla MRI and targeted biopsy of the index prostate cancer: Correlation with robot-assisted radical prostatectomy
.
BJU Int
2017
;
119
:
82
90
.
17.
Cornford
P
,
Bellmunt
J
,
Bolla
M
,
Briers
E
,
De Santis
M
,
Gross
T
, et al
EAU-ESTRO-SIOG Guidelines on Prostate Cancer. Part II: Treatment of relapsing, metastatic, and castration-resistant prostate cancer
.
Eur Urol
2016
.
pii: S0302 - 2838
30469
9
.
18.
Weinreb
JC
,
Barentsz
JO
,
Choyke
PL
,
Cornud
F
,
Haider
MA
,
Macura
KJ
, et al
PI-RADS prostate imaging - reporting and data system: 2015, version 2
.
Eur Urol
2016
;
69
:
16
40
.
19.
Giubilei
G
,
Ponchietti
R
,
Biscioni
S
,
Fanfani
A
,
Ciatto
S
,
F
DIL
, et al
Accuracy of prostate volume measurements using transrectal multiplanar three-dimensional sonography
.
Int J Urol
2005
;
12
:
936
8
.
20.
Kim
S
,
Yu
Z
,
Kil
RM
,
Lee
M
. 
Deep learning of support vector machines with class probability output networks
.
Neural Netw
2015
;
64
:
19
28
.
21.
Balabin
RM
,
Lomakina
EI
. 
Support vector machine regression (LS-SVM)–an alternative to artificial neural networks (ANNs) for the analysis of quantum chemistry data?
Phys Chem Chem Phys
2011
;
13
:
11710
8
.
22.
Carvalhal
GF
,
Smith
DS
,
Mager
DE
,
Ramos
C
,
Catalona
WJ
. 
Digital rectal examination for detecting prostate cancer at prostate specific antigen levels of 4 ng./ml. or less
.
J Urol
1999
;
161
:
835
9
.
23.
Loeb
S
,
Catalona
WJ
. 
What is the role of digital rectal examination in men undergoing serial screening of serum PSA levels?
Nat Clin Pract Urol
2009
;
6
:
68
9
.
24.
Heidenreich
A
,
Bellmunt
J
,
Bolla
M
,
Joniau
S
,
Mason
M
,
Matveev
V
, et al
EAU guidelines on prostate cancer. Part 1: Screening, diagnosis, and treatment of clinically localised disease
.
Eur Urol
2011
;
59
:
61
71
.
25.
Siddiqui
MM
,
Rais-Bahrami
S
,
Turkbey
B
,
George
AK
,
Rothwax
J
,
Shakir
N
, et al
Comparison of MR/ultrasound fusion-guided biopsy with ultrasound-guided biopsy for the diagnosis of prostate cancer
.
JAMA
2015
;
313
:
390
7
.
26.
Ahmed
HU
,
Kirkham
A
,
Arya
M
,
Illing
R
,
Freeman
A
,
Allen
C
, et al
Is it time to consider a role for MRI before prostate biopsy?
Nat Rev Clin Oncol
2009
;
6
:
197
206
.
27.
Moore
CM
,
Robertson
NL
,
Arsanious
N
,
Middleton
T
,
Villers
A
,
Klotz
L
, et al
Image-guided prostate biopsy using magnetic resonance imaging-derived targets: A systematic review
.
Eur Urol
2013
;
63
:
125
40
.
28.
Grey
AD
,
Chana
MS
,
Popert
R
,
Wolfe
K
,
Liyanage
SH
,
Acher
PL
. 
Diagnostic accuracy of magnetic resonance imaging (MRI) prostate imaging reporting and data system (PI-RADS) scoring in a transperineal prostate biopsy setting
.
BJU Int
2015
;
115
:
728
35
.
29.
Hamoen
EH
,
de Rooij
M
,
Witjes
JA
,
Barentsz
JO
,
Rovers
MM
. 
Use of the prostate imaging reporting and data system (PI-RADS) for prostate cancer detection with multiparametric magnetic resonance imaging: A diagnostic meta-analysis
.
Eur Urol
2015
;
67
:
1112
21
.
30.
Park
SY
,
Jung
DC
,
Oh
YT
,
Cho
NH
,
Choi
YD
,
Rha
KH
, et al
Prostate Cancer: PI-RADS version 2 helps preoperatively predict clinically significant cancers
.
Radiology
2016
;
280
:
108
16
.

Supplementary data