Development and internal validation of PI-RADs v2-based model for clinically significant prostate cancer

Prostate cancer (PCa) ranks as the second most common malignancy in male population and has been the second leading cause of cancer-related mortality in Western men [1]. Though the high morbidity and mortality exist, advancements in the early diagnosis attribute much to the improvement of life expectancy. The conventional screening pathway mainly emphasized elevated prostate-specific antigen (PSA) and abnormal digital rectal examination (DRE). However, both the sensitivity and specificity were found to be suboptimal and insufficient for early detection [2].

Multiparametric magnetic resonance imaging (mpMRI) enjoys priority in visualization of prostate due to its high soft-tissue contrast, high resolution, and simultaneous image functional parameters [3]. To set standardized reporting and propose criteria for interpreting data of mpMRI, the European Society of Urogenital Radiology (ESUR) published a reporting system termed Prostate Imaging Reporting and Data System version 1 (PI-RADs v1) in 2012, which was based on four MRI sequences (T2-weighted imaging (T2WI), diffusion-weighted imaging (DWI), dynamic contrast enhanced MRI (DCE-MRI), and MR spectroscopy) [4]. Though PI-RADS v1 system has been validated the accuracy and reproducibility, however, it was not specified exactly how to combine each MRI sequence to derive an overall category assessment, which resulted in confusion in its application. To address this issue, the ESUR and American College of Radiology agreed on the improved PI-RADS version 2 (PI-RADSv2) released online in 2014 [5]. The intended clinical application of PI-RADS v2 is for the diagnostic evaluation as well as risk assessment, and the assessment category of transition zone lesions is mainly determined by the T2WI score while that of peripheral zone lesions is defined by the DWI score [6]. Several studies have validated the high sensitivity and specificity of PI-RADs v2 in diagnosis of prostate cancer [4‐7], and updated PI-RADSv2 shows significant improvement compared with the original Prostate Imaging Reporting and Data System (PIRADS) v1.

There were several risk calculators for PCa, such as European Randomized Study for Screening of Prostate Cancer Risk Calculator (ERSPC-RC), Prostate Cancer Prevention Trial Risk Calculator (PCPT-RC), and Chinese Prostate Cancer Consortium Risk Calculator (CPCC-RC) [8]. The validity of all of the above has been validated in previous studies. However, none of them is composed of PI-RADs v2. The primary objective of this study is to build a model based on PI-RADs v2 and assess its accuracy by internal validation.

It is showed in our study that PI-RADs v2 performed a higher sensitivity and negative predictive value when assessing the detection of csPCa than PCa. And the validation provided evidence supporting both models 1 and 2 that were based on PI-RADs v2, age, and PSAD*10 in predicting csPCa and PCa. The performance of the two models was significantly better than each single variable. Calibration properties were good in patients with PCa and csPCa. These findings were further supported by a decision curve analysis. Several recent studies focusing on the validity of PI-RADs v2 scoring system in detection of csPCa or PCa have validated the diagnostic performance. Though the outcome varied among studies, PI-RADs v2 was proved to have high accuracy for predicting csPCa [2‐4, 6, 7]. One of these studies [6] resulted in AUC of PI-RADs v2-only of 0.83 in PCa and 0.91 in csPCa, which was higher than ours. A possible reason for this might be that our AUC analysis was based on pathological results and experimental examinations, while they made analysis basing on lesions. In their study, patients with suspicious findings, at least one lesion with a PI-RADS v1 assessment category of ≥3, were selected for biopsy and included in the cohort. And that made a great difference. Besides, targeted in-bore MR-guided biopsy helped find more csPCa comparing to our TRUS-guided systematic biopsy of 24-needle cores.

Another study [10] has shown the accurate prediction of PI-RADs v2 based model for high-grade PCa, which also comprised PI-RADs v2, age, and PSAD. On comparing that work to the present study, the model in the present study enrolled more patients (441 versus 247) and showed a lower AUC (83 versus 86%). The reason for this might be that their biopsy was based on targeted lesions whose PI-RADS v1 sum score > 9, and this led to high detection of csPCa. Clinically significant PCa in the present study was defined as GS ≥4 + 3 or 3 + 4 with PSA > 10 ng/ml, > 3 biopsy cores positive, or at least 1 biopsy core with > 50% involvement. Comparing to definition of GS ≥ 7 in their study, less csPCa were observed in our cohort.

There are several predicting tools that have been increasingly developed and validated for use in the PCa screening, such as the European Randomized Study for Screening of Prostate Cancer Risk Calculator (ERSPC-RC) and Prostate Cancer Prevention Trial Risk Calculator (PCPT-RC). Though some variables were found, they were mainly based on age, family history, PSA level, DRE, PV, and previous biopsy status [11, 12]. The Chinese Prostate Cancer Consortium Risk Calculator (CPCC-RC) performed better in decision making of prostate biopsy in Chinese or in other Asian populations included PSA, PV, age, free PSA ration, and DRE but did not involve family history or prior biopsy [8]. However, all the risk calculators above did not take the weight of mpMRI into account. The model established in this study highlighted the dominance of PI-RADS v2 scoring in prediction and showed an AUC of 0.845 (0.786–0.904) for PCa and 0.834 (0.787–0.882) for csPCa in validation cohort, which outperformed the CPCC-RC (AUC 0.801 and 0.826).

The relationship between PSA screening and PCa have been evaluated in both Chinese and Western populations, though it differs importantly between them [13, 14]. A previous study [1] carried out a comprehensive epidemiological analysis of global PCa incidence and mortality using high-quality data. China has the increasing incidence and staple mortality compared to western countries. Prostate volume was proved to be higher in Chinese compared to western population, which could theoretically lead to a higher PSA value and miss PCa at biopsy [14]. PSAD, which could eliminate the influence of PV on PSA, was proved to be a significant predictor for PI-RADs 3–5 lesions [15, 16]. Also, a recent study [17] has validated the incremental value of PSAD in combination with PI-RADS for the accuracy of PCa screening and showed that the NPV of PI-RADS could be improved by inclusion of PSAD and unnecessary biopsies could be reduced. Even for PCa men on active surveillance, combining PSAD and PIRADS score could predict upstaging when PIRADS score is > = 3 with PSAD > 0.15 [18]. We entered PSAD into the model, and it resulted in an excellent diagnostic performance.

In view of the fact that the benefit of mpMRI is becoming an increasingly important aspect of urologic practice [19], there are several reasons that the development of this model should be favored. First of all, it combines PI-RADs v2 with clinical factors PSAD and age, resulting in good clinical performance among both urologists and radiologists. Though moderate inconsistence still exists among the interobserver agreements, PI-RADs v2 reduce variability in imaging by establishing guidelines, summarizing suspicion levels, and standardizing reports. Clinical urologists could improve the diagnostic ability by learning the diagnostic process of PI-RADS v2. Secondly, all patients included in the study received 24-core systematic TRUS-guided biopsy, and the impact on tumor detection of different biopsy methods could be avoided. TRUS-guided systematic biopsy was validated to have similar overall detection compared to MRI-targeted Biopsy or MRI-TRUS fusion biopsy [20, 21], though the detection rate of csPCa might be lower. Last but not the least, this model included only three variables and made it simplified and applied for not only urologists but also radiologists, which was different from previous models.

There are several limitations of this study that should be noticed. The main limitation is a retrospective single-center design, and prospective multicenter external validation should be required to validate its accuracy better. Besides, our outcomes were got according to biopsy-proven Gleason score but not post-prostatectomy pathological grading, which may result in a lower diagnosis quantity of csPCa and make the predictive accuracy of the model be underestimated [22, 23]. Furthermore, we did not enter DRE which was previously proved even a better predictor than PSA into the model, because we wanted the model as objective as possible. And DRE was often performed by resident physicians in our center, which led to a wide difference when it came to the results positive or negative.

We recommend a further study on how would the model performed if we take PI-RADS v2 score 3 as the threshold rather than 4 in current study. And whether this model could be used to assess the diagnostic concordance of csPCa between biopsy results and post-prostatectomy pathological results will be explored in our next study.

The model based on age, PSAD, and PI-RADs v2 score showed internally validated high predictive value for both PCa and csPCa. It could be used to improve the diagnostic performance of suspicious PCa. However, further multicenter external validation should be performed for its wide application.