To assess the role of radiomic features from high b-value DWI sequences in the early detection of clinically significant prostate cancer (PCa)

Purpose or Learning Objective To assess the role of radiomic features from high b-value DWI sequences in the early detection of clinically significant prostate cancer (PCa). Methods or Background 76 patients are retrospectively enrolled, who undergo multiparametric MRI (mpMRI) and biopsy examination, where they received a Gleason Score (GS)=3+3 representing non-clinically significant PCa (ncsPCa, n=26) or GS≥3+4 meaning clinically significant PCa (csPCa, n=50). PCa Regions of Interest (ROIs) are outlined on DWI at b=2000s/mm2 and eighty-four local first-order radiomic features are extracted. First, the LASSO-based method selects a subset of relevant features, discarding linearly correlated couples. Then, to prevent overfitting, only the couple with the lowest p-value at Wilcoxon rank-sum test is selected. A Support Vector Machine (SVM) is trained on 48 patients, validated through 3-fold Cross Validation and tested on 28 patients. ROC curve and AUC are used to assess the SVM performance, together with specificity, sensitivity, and Positive Predictive Value (PPV). Results or Findings The AUC of the ROC curve on the training set is 0.86, with specificity and sensitivity equal to 94% and 77%, respectively, whilst the AUC on the test set is 0.84 with specificity and of 75% and 90%, respectively, and PPV=90%. ncsPCa and csPCa are separated with p=0.007. Conclusion The classifier shows a very low probability of overtreatment of ncsPCa while a high PPV strongly improves the performances of clinical mpMRI used in triage pre-biopsy setting. Limitations This study has been carried out on a 3T machine.