In this study, we develop a model for predicting pECE+ based on only four characteristics: prostate biopsy GS, two classical semantic features on MRI (measurable ECE on MRI and capsular disruption), and TCCL. This model is accurate and reproducible, with goodness-of-fit proved by its AUC, which is 85% for the validation group and approximately 90% for the test group.
First, we evaluated the impact of each classical semantic MRI feature proposed by the ESUR for predicting pECE+, which can be categorized into early signs of ECE: capsular irregularity, capsular bulging, and unsharp margin, and late signs of ECE: obliteration of the rectoprostatic angle, invasion of periprostatic fat, and measurable ECE on MRI. These late signs of ECE are very uncommon in pECE− patients, as we have proven and as reported by other similar studies [12, 13, 18]. These late signs of ECE indicate the presence of significantly invasive clusters of neoplastic cells, which produce irregular disruption of the prostate capsule, and subsequently, infiltration into periprostatic fat, which is observed in advanced PCa stages [12]. In our study, almost all patients with measurable ECE were pECE+, excluding only two patients (1.6%). On reviewing these two false-positive cases, we concluded that in one case, there was a hemorrhage hampering the interpretation of the images, and in the other, there was granulomatous prostatitis coexisting with the PCa. When measurable ECE is identified on MRI, the radiologist can report the presence of pECE with a high degree of confidence. However, measurable ECE is a relatively late marker of pECE and becomes visible predominantly in the advanced stages of PCa. Therefore, it should be reckoned that its absence does not rule out pECE. At our institution, the majority of pECE+ cases operated on had minimal (< 5 mm) pathologic periprostatic extension. This explains why measurable ECE is observed in only 45% of all pECE+ cases.
The critical feature in our model is the introduction of TCCL and a prostate biopsy GS of ≥7 (4 + 3) as significant predictors of pECE+ in unadvanced cases. We proved that TCCL is an independent and reproducible predictor of pECE+, which corroborates a recent review conducted by Kim et al. [19]. The optimal cutoff value for predicting pECE has not yet been established by previous studies and varies between 10 mm and 20 mm [19]. We posit that the global assessment of TCCL and its integration into the model and the other covariables of the model is the key for obtaining the individual probability of being pECE+. Figure 4 presents a demonstration of the probability of being pECE+ calculated using our predictive model. Measurable ECE on MRI is the strongest predictor of pECE, with an approximately 80% probability of the patient being pECE+ when present alone. TCCL and prostate biopsy GS take on a significantly relevant role when measurable ECE and capsular disruption are absent (Fig. 4: black and green lines). For example, in patients with an aggressive prostate biopsy GS, with no measurable ECE or capsular disruption observed on MRI, and a TCCL of > 25 mm, the probability of being pECE+ is greater than 50%. With the same imaging features, but with a marginally aggressive prostate biopsy GS, the TCCL cutoff would need to be increased to approximately 40 mm to achieve a similar probability of being pECE+, i.e., > 50%. Only [18] developed a predictive grading system for predicting pEPE called MRI-derived EPE (MRI-EPE) grade, which combines MRI semantic imaging features and TCCL into three grades. The MRI-EPE grade 1 corresponds to TCCL > 1.5 cm, with a 24% risk of being pEPE+. This score does not account for the influence of the GS and MRI images on the final prediction, which makes it dissimilar from our estimation model. We prove that the same TCCL value (> 1.5 cm) corresponds to different risk levels of being pEPE+ depending on the aggressivity of the GS. For example, in our estimation model, the estimated probability of being pECE+ for a patient with a TCCL of 15 mm is approximately 10% with a marginally aggressive GS, and the probability increases to 30% for patients with a fully aggressive GS. In contrast, [18] did not assess the inter-reader variability of their MRI grading system.
With these predictors (GS and TCCL), our model can diagnose more patients with pECE+ who have no measurable ECE on MRI in the early stages of PCa, improving the global sensitivity to 86% and maintaining moderate-to-high specificity at approximately 73%. These values are in agreement with a recent meta-analysis by Kim et al., which reports a sensitivity and specificity of 0.79 and 0.67, respectively, and an AUC of 0.81 [19]. However, this differs slightly from the meta-analysis by De Rooij et al., which reports a sensitivity and specificity of 0.57 and 0.91, respectively, meaning that the MRI scan had a high specificity but poor and heterogeneous sensitivity for local PCa staging [10]. The differences in these results are attributable to the use of objective parameters such as TCCL in combination with subjective semantic features to predict pECE+ in the studies analyzed by Kim et al. contrasted with the studies included in the meta-analysis by De Rooij et al., which used only subjective semantic features to predict pECE+. The features in the former may have crucial clinical implications, as a high sensitivity or highly specific interpretation may be preferred, depending on the clinical scenario. For example, high sensitivity is required when selecting patients for enrollment in active surveillance programs or choosing candidates for RP with neurovascular bundle sparing. On the other hand, high specificity could be favored when the objective is to avoid potential curative treatment delays. This model shows a good performance to select patients in the early stages of the disease who are the best candidates for RP with neurovascular bundle sparing reducing the side effects conditioned by more invasive surgery. Accurate pre-operative of pECE can also change the therapy decision leading to the recommendation of adjuvant radiotherapy after RP.
Consistent with the literature, our model confirms that the combination of MRI features with a pre-treatment biopsy GS is superior to imaging features alone for predicting pECE [20]. The GS improves the diagnostic accuracy regarding the aggressivity of the disease, and it has been incorporated into many predictive nomograms for detecting ECE [14, 21].
The clinical covariates of age and index lesion PI-RADS score were not associated with pECE (p-values > 0.10), as reported in previous studies [18].
The covariate, PSA levels, was also not significant, even when categorized in the following groups: < 10, 10–20, and > 20. In our study, PSA was not selected for predicting pECE, which is dissimilar from the specifications of existing classical cancer nomograms extensively used in the literature, such as Cancer of the Prostate Risk Assessment (CAPRA), Partin tables, and the Memorial Sloan Kettering Cancer Center (MSKCC) nomogram [21,22,23]. PSA is also used in more recent nomograms based on MRI and Briganti nomograms as a preoperative parameter influencing ECE prediction [24]. The results of our study might be related to the small number of patients with high PSA levels in our sample data: only four patients had PSA levels over 20 ng/ml, which is insufficient for detecting statistical differences between pECE+ patients and pECE− patients.
Our model presents good agreement between the MRI readers for the presence of TCCL and measurable ECE (ICC of 0.82 and 0.88, respectively). However, there was insufficient agreement regarding capsular disruption. The other imaging covariables are extremely heterogeneous between readers and were not significant in the multivariable model.
Our study has some limitations. First, the study sample is relatively small, and the study was performed at a single center using the same MRI protocol and a 3 T MRI scanner. To overcome this bias, we introduced a validation sample of 59 patients whose MRI examinations were performed outside the primary study facility and with different MRI equipment to increase the robustness of the model. We did not identify statistical differences between the test and validation groups (Table 4) for the variables under consideration. Nevertheless, this result does not allow to conclude that the heterogeneity between MRI’s acquisitions protocol and technical specifications of the equipment in the validation sample could modify interpretation accuracy. Furthermore, there is a sampling bias resulting from the selection of prostatectomy specimens as the histopathological reference standard because prostatectomy was not proposed for the significantly advanced cases.
Second, we focused only on the index lesion identified on MRI, and it was the only one correlated with the prostate lesion. We did not take tumor multifocality on MRI and pathology into account.
Third, the TCCL measurements presented in our study may be limited by our institutional software and sequence specifications and require validation at other institutions and the use of other MRI protocols.
Our model uses measurable ECE as a determinant MRI semantic feature to detect pECE. Although measurable ECE is correlated between readers, it has a low prevalence in pECE+ patients because it is observed predominantly in significantly advanced cases. The other considerably prevalent semantic feature (capsular disruption) is not significant for detecting early-stage cases of pECE and is not so correlated between readers. Hence, we need to conduct additional research on preoperatively detecting microscopic ECE using a grading of objective markers, such as TCCL and GS, and validate these markers at various institutions or incorporate a new artificial intelligence (AI) analysis into the estimation model developed in this study. A computer-based method of extracting and analyzing image features qualitatively (Radiomics) could provide more information about the PCa tumor facilitating risk stratification and therapeutic management of these patients. MRI has been the most studied imaging modality for radiomics application in PCa, so far, but more research is warranted in order to get robustness of MRI- based radiomics features models [25].