Local staging with multiparametric MRI in daily clinical practice: diagnostic accuracy and evaluation of a radiologic learning curve

Prostate cancer (PCa) is the most common cancer in men of older age in Western countries [1]. Accurate staging of the primary tumor is of vital importance, as the distinction between organ-confined disease (T-stage 1–2) versus locally advanced tumors (T3–4) influences both prognosis [2] and treatment planning [3].

The main therapeutic approaches for PCa include radical prostatectomy and radiotherapy [3]. When considering a radical prostatectomy, the presence of locally advanced disease warrants a concomitant extended pelvic lymph node dissection (ePLND), as there is an increased risk of lymph node metastasis [3‐5]. In addition, local tumor stage guides surgical planning regarding the preservation of the neurovascular bundle. Nerve-sparing surgery is generally restricted to patients with organ-confined disease. Extension of PCa outside the prostatic capsule requires dissection of the neurovascular bundle, for nerve-sparing surgery would increase the risk of positive surgical margins [3, 4, 6]. The assessment of the local tumor stage is similarly important when radiotherapy is chosen as treatment and guides decisions on radiation dose, radiation template, and adjuvant therapies [3, 7].

For assessment of the local tumor stage, routine diagnostics (i.e., digital rectal examination, serum prostate-specific antigen (PSA) level, transrectal ultrasound, and biopsy Gleason score [3, 8]) are insufficient [9]. When combining these clinical parameters into predictive nomograms (e.g., the Partin Tables), staging accuracy increases, but remains imperfect [10, 11].

To overcome this diagnostic shortcoming, multiparametric magnetic resonance imaging (mpMRI) is increasingly deployed. mpMRI is an imaging technique that combines different (functional) imaging sequences, generating improved detection, and localization of malignant lesions. Although mpMRI presents promising detection of PCa [12], accurate assessment of the tumor stage is still imperfect. In a recent meta-analysis, the sensitivity of mpMRI for overall T3 detection reached 61% only (95% CI 54–67%) [13].

A concern regarding mpMRI is the considerable inter-observer variability [14, 15]. This problem might be due to different experience of radiologists with mpMRI, as a marked radiologic learning curve was demonstrated [16‐19]. The presence of such learning curve, however, is studied mainly for primary detection of PCa. Research specifically evaluating the existence of a radiological learning curve for correct staging of PCa is scare, focusing mainly on endorectal MRI [18]. In the cited meta-analysis, the effect of radiologists’ experience on diagnostic accuracy was evaluated, but the results were inconclusive [13].

In this study, we aimed to estimate the diagnostic accuracy of mpMRI for the detection of locally advanced PCa stages (pT3–4) prior to radical prostatectomy, in a multicenter, real-life clinical cohort of patients. We additionally assessed the diagnostic accuracy over time, evaluating the existence of a radiologic learning curve.

An overview of the patients’ characteristics is presented in Table 1. A PI-RADS classification was given in 60.0% of all cases (rising from 0% in 2012 to 65.1% in 2016). Pathology analysis following radical prostatectomy revealed extra-prostatic extension (pT3a) in 76 (18%) patients, seminal vesical invasion in 57 (13%) patients, and advancement of the tumor into adjacent structures (pT4) in 4 (1%) patients.

The presence of malignancy was correctly detected by mpMRI in n = 358 patients (sensitivity 84.0%, CI 80.6–87.5%). In Table 2, the findings on preoperative mpMRI (rT) and concurrent RARP pathology results (pT) are depicted.

mpMRI detected 62 out of 137 patients with locally advanced disease (pT3–4), resulting in a sensitivity of 45.3% (CI 36.9–53.6%), specificity 75.8% (CI 70.9–80.7%), PPV 46.6% (CI 38.1–55.1%), and NPV of 74.7% (CI 69.8–79.7%). Sensitivity in the group with high risk of locally advanced disease (n = 133) was 49.2% (CI 36.4–61.9%) versus 42.3% (CI 31.4–53.3%) in the lower risk group (n = 297) (p = 0.49). Specificity was 73.0% (CI 62.9–83.1%) and 76.3% (CI 70.6–82.0), respectively.

Radiologic understaging (i.e., the failure to detect locally advanced disease) occurred in n = 75 cases (54.7% of all patients with pT3–4). In n = 30 of these patients, no ePLKD was performed (21.9% all patients with pT3–4), and in n = 41 patients, complete nerve preservation was performed (29.9% of all pT3–4). This reveals the potential undertreatment associated with incorrect radiologic staging. Radiologic overstaging (incorrect detection of locally advanced disease) was present in n = 70 cases (23.9% of patients with pT0–2). In n = 16 of these cases, no form of nerve-sparing was performed (5.5% of all patients with organ-confined disease), revealing potential overtreatment.

In Fig. 1, the diagnostic accuracy of mpMRI for the detection of locally advanced disease is presented for all study years separately (radiologic learning curve). Over time, a negative trend was observed, although the differences in diagnostic accuracy were not statistically significant (sensitivity per year p = 0.12, specificity per year p = 0.57; sensitivity per half sample p = 0.61).

The preoperative assessment of local tumor stage affects important therapeutic decisions regarding radicality of the surgical procedure (i.e., a wide or less wide excision around the prostate; the performance of an ePLND) and the ability to preserve structures that relate to functional outcomes (i.e., to perform nerve-sparing surgery or not).

In our extensive, multicenter cohort, the accuracy of mpMRI for local tumor staging was evaluated. The reference standard was the pathology report following RARP. Intriguingly, more than half of the cases (55%) with locally advanced disease (pT3–4) remained undetected by preoperative mpMRI (see Table 2). Conversely, if mpMRI indicated locally advanced disease, in more than half of the cases, these results proved to be falsely positive, as pathological examination showed organ-confined disease. We observed no significant increase in diagnostic accuracy when confining the use of mpMRI to patients with high risk of locally advanced disease only.

Our results are in line with those observed in the review by De Rooij et al. [13]; and taking together, it appears questionable whether mpMRI is adequate to guide therapeutic decision-making. Moreover, mpMRI does not seem superior in predicting locally advanced disease compared to routine clinical parameters such as those used in the Partin tables [11]. There may be a role for mpMRI when results are combined with clinical parameters, as a recent study showed some increased accuracy for such combination (not a predefined nomogram) [23]. Future research might clarify if the incorporation of mpMRI to the established nomograms could increase diagnostic confidence.

Novel insights of this study include the clinical impact of erroneous staging and the evaluation of a radiologic learning effect. Radiologic understaging potentially impairs oncologic outcomes, as patients might not receive the indicated ePLND or might be operated with a possibly hazardous nerve-sparing approach. We have shown that in a third of all patients with locally advanced disease, radiologic understaging might have contributed to a form of undertreatment.

Radiologic overstaging, on the other hand, leads to patients unnecessarily being withheld nerve-sparing surgery, causing impaired potency and urinary continence. In our cohort, for 16 men with organ-confined disease, this form of overtreatment seemed present (6% of all patients with organ-confined disease).

The above-mentioned numbers should be interpreted cautiously. Due to the retrospective nature of our study, causality cannot be proven. Furthermore, no complete insight into surgical decision-making was present. For example, the choice to not perform an ePLND could have been caused by radiologic understaging, but patient factors outside the scope of this study may also have been the reason to omit an ePLND (e.g., a technically unfeasible procedure). The true extent of undertreatment, caused by inaccurate mpMRI findings, might, therefore, have been more limited.

The same applies the identified patients with potential overtreatment. We cannot know whether all of these patients would have opted for a nerve-sparing approach, if preoperative radiologic staging would have been correct. Preoperative erectile function, for example, may have been absent, rendering nerve-sparing surgery futile.

In our cohort, no radiological learning curve could be observed. Despite the rising experience with mpMRI and increasing number of performed procedures, the sensitivity for locally advanced disease did not improve over the years. In the previous reports as well, the number of included procedures had no influence on diagnostic accuracy [13], challenging the presumed gain of concentrating mpMRIs to a limited number of hospitals.

Although our sample size was substantial (430 patients), one might question whether 137 patients with locally advanced disease divided over different hospitals, over multiple years’ time, were sufficient to facilitate learning. It is important to realize, however, that radiologists were exposed to more mpMRI studies than included in this analysis (i.e., also radiotherapy patients are staged with mpMRI). Besides, the included centers are all recognized reference centers for PCa care, meaning that these numbers comprise the realistic clinical volumes in the present-day oncologic care.

In this study, we examined the existence of a radiologic learning curve on a hospital level. We cannot formally exclude the possibility that a learning curve was present for individual radiologists, as no strict, prospective protocol was followed. However, it is frequently stated that mpMRI should be confined to expert centers [3, 24], implying general hospital-level learning. As individual radiologists may change employment, the evaluation of a learning curve irrespective of changes in individual expertise is warranted.

Our study has some additional limitations. The available information on the radiologists performing mpMRI interpretation was limited. In the included hospitals, prostate mpMRIs are strictly reviewed by dedicated radiologists. Although this is widespread practice, we cannot formally exclude the possibility that less experienced radiologists have occasionally reviewed scans in referring centers. Standardly performing a central revision of the mpMRI images might have overcome such imperfection, but is not in line with the current clinical practice.

Another point is the impact of the rising number of scans on clinical workflow. We hypothesized that such increase would lead to more radiologic experience and thereby to higher diagnostic accuracy. However, we do not know that the available time radiologists were given to review all scans. Given the marked rise in number of scans, the available time per scan could have been comprised, possibly offsetting a potential learning effect.

The strength of our analysis lies in the fair representation of clinical reality in the present-day oncologic centers. If the volumes, methods, and interpretation of mpMRI acquisitions in these acknowledged reference centers do not lead to accurate results, it seems challenging to ensure widespread value of the imaging technique.

In this multicenter, real-life clinical cohort of patients, the diagnostic accuracy of mpMRI for the detection of locally advanced PCa stages was limited. Therefore, it remains questionable whether the diagnostic accuracy of mpMRI is sufficient to guide therapeutic decision-making. In addition, no significant radiologic learning curve for the detection of locally advance disease was observed.