Session Item

Prostate cancer (PCa) is the 2nd most common cancer in men. Several treatment options are available, making an accurate diagnostic and risk stratification essential to maximize the therapeutic benefits. Several oncological and imaging-based scores are conventionally used for PCa characterization, such as the Gleason Score (GS), National Comprehensive Cancer Network risk class (NCCN risk class), T-stage, extracapsular extension (ECE) score, and Prostate Imaging Reporting and Data System v2 (PIRADS). In this context, multiparametric MRI is showing promising results for extracting information sensitive to pathological differences in PCa, providing useful tools for non-invasive tumour characterization. The aim of this study is to investigate the potential of diffusion MRI-based radiomic models as a non-invasive tool for Gleason Score prediction in PCa patients treated with radiotherapy.

65 PCa patients, who underwent diffusion-weighted MRI (DWI) and enrolled for radiotherapy between 2014 and 2018 at the European Institute of Oncology (IEO, Italy), were included in the study. From mono-exponential fits of DWI, apparent diffusion coefficient (ADC) maps were estimated for 50 eligible patients. 107 radiomics features (14 shape, 18 first-order, 75 texture) were extracted from ADC maps of the whole prostate glands. A comparison was made of 5 feature selection routines (correlation, mutual information, Relief, RFECV, Mann-Whitney U-test), with the most predictive features from each being fed to three different classification models (logistic regression (LR), support vector machine (SVM), random forest (RF)) investigating the ability to predict total GS (6 (52%) vs. 7 (48%)) starting from ADC.  Classification models were encapsulated in a 5-fold cross-validation routine. In addition, receiver operating characteristic (ROC) curves were built, and the average precision (AP) was calculated to access the accuracy of classification models. The predictive powers of the models built with radiomics features were finally compared, in terms of F1-score, with the ability of conventional clinical scores (i.e., NCCN risk class, T-stage, ECE, PIRADS)  to stratify patients according to total GS.

From AP and ROC analyses, the best diagnostic performance was found using RFECV as feature selection and LR as classifier, reaching an AP of 0.78 and an area under the ROC curve (AUC) of 0.81 (Fig1) . Among the 36 features with the highest predictive performance for GS, the textural ones were found to be the most frequent (12 shape, 6 first order, 18 texture). In the F1-score analyses, the radiomics-based model was found to be more powerful in predicting GS (F1=0.71, Fig.1) than the conventional clinical scores (F1=0.42, 0.19, 0.19, 0.12 for ECE, NCCN risk-class, PIRADS, and T-stage, respectively).

Radiomic models based on DWI are a promising non-invasive tool for PCa characterization implying advantages for personalized therapy approaches.