Beyond Multiparametric MRI and towards Radiomics to Detect Prostate Cancer: A Machine Learning Model to Predict Clinically Significant Lesions

SIMPLE SUMMARY: Early diagnosing clinically significant prostate cancer (csPCa) through Magnetic Resonance Imaging (MRI) is very challenging and, nowadays, csPCa confirmation comes exclusively from prostate biopsy. However, biopsy is an invasive procedure and it also frequently causes csPCa misclassification. This study develops a non-invasive machine learning method for early predicting csPCa based on radiomic MRI image analysis. The main novelty of this study is investigating the radiomic differences between progressive risk groups of PCa, attributed according to the biopsy outcome and Gleason Grade group stratification. Besides predicting csPCa with very good performance (sensitivity = specificity = 84% in the test phase), this study highlights patients with GG = 2 as non-csPCa (being statistically equivalent to GG = 1), so that GG = 2 can be counselled for follow up, whilst GG ≥ 3 admitted to radical treatments. Not least, this study provides a plausible clinical interpretation of radiomic features, by discussing their values with respect to the histological meaning. ABSTRACT: The risk of misclassifying clinically significant prostate cancer (csPCa) by multiparametric magnetic resonance imaging is consistent, also using the updated PIRADS score and although different definitions of csPCa, patients with Gleason Grade group (GG) ≥ 3 have a significantly worse prognosis. This study aims to develop a machine learning model predicting csPCa (i.e., any GG ≥ 3 lesion at target biopsy) by mpMRI radiomic features and analyzing similarities between GG groups. One hundred and two patients with 117 PIRADS ≥ 3 lesions at mpMRI underwent target+systematic biopsy, providing histologic diagnosis of PCa, 61 GG < 3 and 56 GG ≥ 3. Features were generated locally from an apparent diffusion coefficient and selected, using the LASSO method and Wilcoxon rank-sum test (p < 0.001), to achieve only four features. After data augmentation, the features were exploited to train a support vector machine classifier, subsequently validated on a test set. To assess the results, Kruskal–Wallis and Wilcoxon rank-sum tests (p < 0.001) and receiver operating characteristic (ROC)-related metrics were used. GG1 and GG2 were equivalent (p = 0.26), whilst clear separations between either GG[1,2] and GG ≥ 3 exist (p < 10 [Formula: see text]). On the test set, the area under the curve = 0.88 (95% CI, 0.68–0.94), with positive and negative predictive values being 84%. The features retain a histological interpretation. Our model hints at GG2 being much more similar to GG1 than GG ≥ 3.