An Apparent Diffusion Coefficient-Based Machine Learning Model Can Improve Prostate Cancer Detection in the Grey Area of the Prostate Imaging Reporting and Data System Category 3: A Single-Centre Experience

SIMPLE SUMMARY: Multiparametric Magnetic Resonance Imaging (mpMRI) interpretation and reporting is based on the more recent version 2.1 of the Prostate Imaging-Reporting and Data System (PI-RADS), revised in 2019, indicating the probability of clinically significant Prostate Cancer (csPCa) on a 5-point scale, which should be confirmed through trans-rectal ultrasound (TRUS) fusion-targeted biopsy. Among PI-RADS categories, PI-RADS 3 lesions represent a highly “equivocal” result, with a non-negligible probability of PCa, or even csPCa. This study exploits machine learning methods in order to investigate the role of mpMRI as a stand-alone tool for early and non-invasive detection of PCa in a selected cohort of PI-RADS 3 lesions, by means of a radiomic analysis of Apparent Diffusion Coefficient sequences. Differently from what reported in the current literature, the methodology adopted has bounded the possibility of overoptimistic predictive performance, also improving the state-of-art by achieving a positive predictive value of 80%, with specificity = 76% and sensitivity = 78%. ABSTRACT: The Prostate Imaging and Reporting Data System (PI-RADS) has a key role in the management of prostate cancer (PCa). However, the clinical interpretation of PI-RADS 3 score lesions may be challenging and misleading, thus postponing PCa diagnosis to biopsy outcome. Multiparametric magnetic resonance imaging (mpMRI) radiomic analysis may represent a stand-alone noninvasive tool for PCa diagnosis. Hence, this study aims at developing a mpMRI-based radiomic PCa diagnostic model in a cohort of PI-RADS 3 lesions. We enrolled 133 patients with 155 PI-RADS 3 lesions, 84 of which had PCa confirmation by fusion biopsy. Local radiomic features were generated from apparent diffusion coefficient maps, and the four most informative were selected using LASSO, the Wilcoxon rank-sum test (p < 0.001), and support vector machines (SVMs). The selected features where augmented and used to train an SVM classifier, externally validated on a holdout subset. Linear and second-order polynomial kernels were exploited, and their predictive performance compared through receiver operating characteristics (ROC)-related metrics. On the test set, the highest performance, equally for both kernels, was specificity = 76%, sensitivity = 78%, positive predictive value = 80%, and negative predictive value = 74%. Our findings substantially improve radiologist interpretation of PI-RADS 3 lesions and let us advance towards an image-driven PCa diagnosis.