2.6. Statistical Analysis

The endpoint of this study was an investigation of the diagnostic performance of Radiomics features extracted from multimodality imaging (MRI-ADC and [⁶⁸Ga]Ga-PSMA-11 PET) against the ISUP grade obtained from histological evaluation, in particular, the ability of radiomic features to discriminate ISUP 1 from higher grades in order to help with treatment stratification. In Figure 2, the entire process of statistical analysis is summarized, from features extraction to final model evaluations.

Detail of the workflow employed, from features extraction to selection of the final statistical models.

Five independent predictive logistic models for ISUP Grade were developed based on:

lesions visible only through [⁶⁸Ga]Ga-PSMA-11 PET;

lesions visible only with MRI-ADC;

lesions visible with [⁶⁸Ga]Ga-PSMA-11 PET and MRI-ADC but only employing 68-[68Ga]Ga-PSMA-11 PET imaging features;

lesions visible with [⁶⁸Ga]Ga-PSMA-11 PET and MRI-ADC but only employing mp-MRI imaging features;

lesions visible both with [⁶⁸Ga]Ga-PSMA-11 PET and MRI-ADC, with features extracted from both imaging modalities.

The models were built through a stochastic cross-validation process to evaluate their performance.

The modeling process followed the following procedure:

Lesion feature datasets were divided into training (2/3) and test (1/3) sets. Subsequently, a logistic regression model was trained on the training set, with the employing features selected by a least absolute shrinkage and selection operator (LASSO) algorithm with internal 3-fold cross-validation. The predictive ability of the model was then calculated on the test set. This operation was repeated 30 times and subsequently, receiver operating curves (ROC) and the area under the curve (AUC) of each iteration were recorded both for the training and test sets.

The models’ quality was reported by averaging the AUC across iterations. The ROC and AUC were reported for the best-performing iteration to evaluate the model’s prediction power and to compare the performances of the mixed imaging features model (e) with standalone imaging models (c) and (d). The most frequently selected features across iterations were reported as the most informative features for ISUP Grade prediction.

All statistical analyses were carried out with R and the open-source software RStudio [41]. The raw data of this study ([⁶⁸Ga]Ga-PSMA-11 PET, MRI-ADC and Pathology records) are available from the corresponding author on reasonable request.