Radiomics features selection and modeling
All CT scans, derived from 78 patients, were considered technically adequate for the purpose of the analysis and therefore included in the study. In the figure 2 is shown the workflow of the study. Images were acquired on two different scanners (SOMATOM Plus 4 before 2011, SOMATOM Definition Flash after 2011, Siemens Healthineers, Erlangen, Germany). Mean pixel spacing was 0.44 (0.42-0.45) while mean slice thickness 2.6 mm (1-5 mm). In total, 232 radiomics features have been firstly extracted, belonging to the following feature classes: 20 statistical features (grey-level histogram); 14 morphological features; 100 texture features GLCM (grey level co-occurrence matrix); 66 texture features GLRLM (grey level run length matrix); 32 texture features GLSZM (grey level size zone matrix). After the Boruta selection procedure, 8 features were selected and addressed to the further step (Figure 3). After Pearson correlation analysis, 2 features were lastly retained: F_stat.mean (mean of voxel intensity histogram) and F_szm_2.5D.zsnu (zone size non-uniformity, computed in the 2.5D version). According to IBSI definition, F_stat.mean is a morphological intensity-based statistical feature that describes the distribution of the grey levels within the considered ROI.F_szm_2.5D.zsnu is a textural feature that assesses the distribution of zone counts over the different zone sizes. The uniformity of the zone sizes is low when the zone counts are distributed equally along zone sizes [16].