Session Item

Radiomic and clinical features combination for the prediction of pathological characteristics of prostate cancer (PCa) may pave the way for improved decision-making and personalization of treatment. It is often not clear however, which radiomic features (RadF) contribute, and how these contributions differ, depending on the pathological feature being modelled. Aim of the study was to compare the contributions of radiomic signatures in the prediction of PCa features with prostatectomy as confirmation.

A representative subset of 100 patients (pts) from the cohort of about 1500 who have undergone MRI and prostatectomy in our Institution since 2015 was selected by balancing the clinical characteristics of the pts. Prostate of each patient was segmented from T2-weighted axial MRI images by an expert radiologist, and 1810 RadF were extracted (PyRadiomics v3.0.1, AIM-Harvard). Radiomic set was reduced to 50 features using a hierarchical clustering procedure based on absolute rank correlation; in each cluster, the feature with the highest absolute rank correlation with the target variable was selected. Gradient-boosted decision-tree models were separately trained using clinical variables (age, prostate volume, iPSA, EPE score and PI-RADS category, biopsy-based total Gleason score and ISUP grade, and risk class) alone and in combination with the selected RadF to predict surgical marginal status (R0 vs R1), pathology-based lymph node status (pN0 vs pN1), tumor stage (pT2 vs pT3) and ISUP grade group (≤3 vs ≥4), and validated with repeated 5-fold cross validation. The mean feature importance in the clinical + radiomic models was determined based on mean prediction value change over validation folds.

Validation AUC values (±95% CI) of the different models were 0.800 (±0.007) for surgical marginal status, 0.871 (±0.010) for pathological lymph nodes, 0.795 (±0.006) for pathological tumor stage, and 0.877 (±0.009) for ISUP grade group (Fig1). The contributions of the top eight RadF in each model are also displayed (Fig2). In the models for pathological lymph node status and tumor stage, both EPE score and PI-RADS category had a large impact on the predictions, while none of the clinical variables appeared in the top eight for prediction of surgical marginal status or pathology-based ISUP grade group. In terms of important RadF, we had Laplacian of Gaussian (“log”) features for surgical marginal status, local binary pattern (“lbp”) features for pathological tumor stage, and wavelet features for ISUP grade group.

Radiomics can have a significant impact on the prediction of the pathological features of PCa; radiomic and clinical features, in particular the radiological features of PI-RADS category and EPE score, appear to complement each other with different roles in the prediction of different pathological features. Different types of RadF can be important in different contexts and should not be judged on an absolute utility scale.