Diagnosis of Clinical Significant Prostate Cancer on Biparametric MRI Using Zone-Specific Radiomic Features

Quantitative assessment of MRI, by means of radiomic analyses, is an emerging approach for prostate cancer (PCa) detection and characterization. Typically, radiomic features are extracted from the lesions, despite inherent uncertainties surrounding PCa segmentation. The aim of the study was to assess the usefulness of mpMRI-based radiomic models, originating from distinct anatomical regions of the prostate for non-invasive characterization of clinically significant PCa and compare them with lesion-derived radiomic models. Different classification tasks were formulated for each anatomical region (whole gland, peripheral zone, transition zone) and the corresponding lesions. For each task, four sets of radiomic features were considered (T2w, DWI, ADC, and their combination), and four classification algorithms (LASSO, RF, SVM, XGB) were implemented. Nested cross-validation was applied for model development, feature selection, hyperparameter optimization, and performance assessment. Whole-region RF radiomic models, with a maximum AUC of 0.84, outperformed the corresponding tumor-specific radiomic models (maximum AUC= 0.75).