Figure 2.

Radiomics’ workflow. (1) 324 GEP NET lesions with high SSTRs expression in [⁶⁸Ga]DOTATOC PET were analyzed (LIFEx) placing a 2D-circular ROI (at least 16 voxels, 0.443 cm³) on the lesion’s part with the highest SUV_max. (2) 65 features from each lesion (parenchyma, lymph nodes, bones) + additional features: Pre-PRRT CgA values and grading (G1-G2-G3) were assessed. (3) Descriptive-inferential sequential approach for feature reduction and selection; for each feature, the point biserial correlation index between the features and the dichotomic variable (0 PD vs. 1 SD, CR, PR) was calculated, sorting the features in descending order. Then, a cycle started to add one column at a time performing a logistic regression analysis comparing the p-value of each iteration and stopping in case of a growing p-value. (4) Discriminant analysis was then used for feature classification using the most discriminative ones identified in the previous step.