Radiomics for predicting extraprostatic extension of prostate cancer with MRI: a systematic review and meta-analysis
Radiomics for predicting extraprostatic extension of prostate cancer with MRI: a systematic review and meta-analysis
Andrea Ponsiglione, Michele Gambardella, Arnaldo Stanzione, Roberta Green, Valeria Cantoni, Carmela Nappi, Felice Crocetto, Renato Cuocolo, Alberto Cuocolo, Massimo Imbriaco
To systematically review the literature and perform a meta-analysis of MRI-based radiomics approaches for extraprostatic extension (EPE) prediction of prostate cancer (PCa).
Multiple databases were systematically searched for radiomics studies on EPE detection up to June 2022. Their methodological quality was assessed according to the Quality Assessment of Diagnostic Accuracy Studies 2 tool (QUADAS-2) and Radiomics Quality Score (RQS). For the purposes of data pooling, in case of studies reporting data for either internal or external test-set, we considered them separately, while if different models where built, the one with the best predictive performance was selected. The predictive accuracy (predicting presence of EPE) was quantified using the AUC for the receiver operating characteristic curve analysis. The overall effect size was estimated using a random-effects model, and statistical heterogeneity was evaluated using the I2 value. A funnel plot was used to investigate publication bias. Subgroup analyses were conducted based on the use of a dedicated test-set, deep learning (DL), scanner number, and the integration of clinical data in the model.
The detailed literature search process is shown in Figure 1. Briefly, 260 initial citations were reduced to 13 relevant articles for a meta-analysis. The methodological quality assessment of risk of bias within eligible studies according to QUADAS-2 is shown in Figure 2. Risk of bias due to patient selection was unclear in six studies because they did not provide adequate information on consecutive or random selection. In the index test domain, the risk of bias was unclear in seven studies due to the lack of preprocessing details, while it was high in three studies who did not test feature robustness. One study had unclear risk of bias for inadequately detailing the reference standard. In five cases, the authors did not clearly report the time between MRI and RP, receiving an unclear risk of bias. As for applicability concerns, one study scored at unclear risk of bias due to insufficient patient selection and reference standard details. As for the methodological quality assessment according to RQS, the total RQS ranged from 0 to 33% of the maximum rating, with a median score of 10/36 (interquartile range 11). The RQS was low especially due to the lack of prospective design (all studies were retrospective) and of comparison with gold standard. All investigations performed discriminations statistics while none of them made their code or data publicly available. The radiomics models for EPE prediction showed an overall pooled AUC = 0.80 (95% CI = 0.74-0.86) (Figure 3). Study heterogeneity was 84.7% (p < .001). In the subgroup analysis, the four studies not using dedicated test-set showed a pooled AUC of 0.89 (95% CI = 0.78-0.99) and heterogeneity of 89.6% (p < .001). The pooled AUC for the remaining studies with a dedicated test-set was 0.78 (95% CI = 0.73-0.82) with 46.4% heterogeneity (p = .038). Three studies incorporated DL into their pipeline, yielding a combined AUC of 0.72 (95% CI = 0.60-0.84) with 68.4% heterogeneity (p = .042). The studies not employing DL achieved a pooled AUC of 0.82 (95% CI = 0.76-0.89) with 83.7% heterogeneity (p < .001). Among nine studies using multiple scanners, the pooled AUC was 0.79 (95% CI = 0.74-0.84) and heterogeneity 32.6% (p = .157), while investigations with single scanners yielded AUC of 0.83 (95% CI = 0.73-0.92) and 89.8% heterogeneity (p < .001). Finally, the five studies in which the best predictive performance was achieved by combined models showed a pooled AUC of 0.76 (95% CI = 0.71-0.82) and 14.8% heterogeneity (p = .320). In studies wherein the best predictive performance was obtained with only-radiomics models, pooled AUC was 0.82 (95% CI = 0.75-0.89) and heterogeneity 84.3% (p < .001).
Figure 1. Literature search and study selection process flowchart.
Figure 2. Methodological quality of the included studies assessed according to the Quality Assessment of Diagnostic Accuracy Studies 2 tool for risk of bias and applicability concerns. The green circle represents the low risk of bias, the yellow circle the unclear risk of bias, and the red circle the high risk of bias.
Figure 3. Forest plot of single studies for the pooled area under the curve (AUC) and 95% CI of extraprostatic extension (EPE) prediction. Horizontal lines represent 95% confidence interval of the point estimates. The diamond means the pooled AUC estimate. ^ internal test-set, * external test-set 1, ° external test-set 2.
Radiomics adds complexity to prostate MRI, potentially enhancing personalized EPE assessment of PCa. However, its role must be brought into context with established tools and more practical alternatives. Technical and diagnostic studies suggest radiomics may redefine EPE prediction alongside radiologist evaluation. Methodologically strong research assessing its diagnostic and therapeutic impact is needed.
