Background: Prostate cancer (PCa) represents a matter at the forefront of healthcare, being divided into clinically significant (csPCa) and indolent PCa based on prognostic and treatment options. Although multi-parametric magnetic resonance imaging (mpMRI) has enabled significant advances, it cannot differentiate between the aforementioned categories; therefore, in order to render the initial diagnosis, invasive procedures such as transrectal prostate biopsy are still necessary. In response to these challenges, artificial intelligence (AI)-based algorithms combined with radiomics features offer the possibility of creating a textural pixel pattern-based surrogate, which has the potential of correlating the medical imagery with the pathological report in a one-to-one manner. Objective: The aim of the present study was to develop a machine learning model that can differentiate indolent from csPCa lesions, as well as individually classifying each nodule into corresponding ISUP grades prior to prostate biopsy, using textural features derived from mpMRI T2WI acquisitions. Materials and Methods: The study was conducted in 154 patients and 201 individual prostatic lesions. All cases were scanned using the same 1.5 Tesla mpMRI machine, employing a standard protocol. Each nodule was manually delineated using the 3D Slicer platform (version 5.2.2) and textural parameters were derived using the PyRadiomics database (version 3.1.0). We compared three machine learning classification models (Random Forest, Support Vector Machine, and Logistic Regression) in full, partial and no correlation settings, in order to differentiate between indolent and csPCa, as well as between ISUP 2 and ISUP 3 lesions. Results: The median age was 65 years (IQR: 61–69), the mean PSA value was 10.27 ng/mL, and 76.61% of the segmented lesions had a PI-RADS score of 4 or higher. Overall, the highest performance was registered for the Random Forest model in the partial correlation setting, differentiating between indolent and csPCa and between ISUP 2 versus ISUP 3 lesions, with accuracies of 88.13% and 82.5%, respectively. When the models were trained on combined clinical data and radiomic signatures, these accuracies increased to 91.11% and 91.39%, respectively. Conclusions: We developed a machine learning decision support tool that accurately predicts the ISUP grade prior to prostate biopsy, based on the textural features extracted from T2 MRI acquisitions.

Background: Prostate cancer (PCa) is the most frequent neoplasia in the male population. According to the International Society of Urological Pathology (ISUP), PCa can be divided into two major groups, based on their prognosis and treatment options. Multiparametric magnetic resonance imaging (mpMRI) holds a central role in PCa assessment; however, it does not have a one-to-one correspondence with the histopathological grading of tumors. Recently, artificial intelligence (AI)-based algorithms and textural analysis, a subdivision of radiomics, have shown potential in bridging this gap. Objectives: We aimed to develop a machine-learning algorithm that predicts the ISUP grade of manually contoured prostate nodules on T2-weighted images and classifies them into clinically significant and indolent ones. Materials and Methods: We included 55 patients with 76 lesions. All patients were examined on the same 1.5 Tesla mpMRI scanner. Each nodule was manually segmented using the open-source 3D Slicer platform, and textural features were extracted using the PyRadiomics (version 3.0.1) library. The software was based on machine-learning classifiers. The accuracy was calculated based on precision, recall, and F1 scores. Results: The median age of the study group was 64 years (IQR 61–68), and the mean PSA value was 11.14 ng/mL. A total of 85.52% of the nodules were graded PI-RADS 4 or higher. Overall, the algorithm classified indolent and clinically significant PCas with an accuracy of 87.2%. Further, when trained to differentiate each ISUP group, the accuracy was 80.3%. Conclusions: We developed an AI-based decision-support system that accurately differentiates between the two PCa prognostic groups using only T2 MRI acquisitions by employing radiomics with a robust machine-learning architecture.