Radiomics Analysis of Tumor Volumes

The delineation of the primary gross tumor volume (GTV) was performed using the RayStation treatment planning system (version 8B SP2; RaySearch Laboratories AB, Stockholm, Sweden) on the initial planning CT (W0). The diagnostic FDG-PET images were manually registered to the CT image using rigid registration in the treatment planning system. To provide consistent delineation for the radiomic analyses, all clinical ROIs were inspected and re-contoured if required by a radio-oncologist in training (K.L.) and validated by an experienced radio-oncologist (E.G.C.T.). For W2 and W3, contours from W0 were transferred and adapted to the GTV based on the corresponding CT imaging data. FDG-PET intensities were converted to body weight-corrected standard uptake values (SUV). Patient images were then interpolated to isotropic voxel spacing (1 × 1 × 1 ${\mathrm{mm}}^3$ and 3 × 3 × 3 ${\mathrm{mm}}^3$ for CT and FDG-PET, respectively) using cubic splines. Segmentation masks were linearly interpolated and binarized using a cutoff of 0.5, and 182 radiomics features capturing statistical, morphological, intensity based and texture based characteristics were computed from each modality at every timepoint using the ‘medical image radiomics processor’ package (MIRP) [36 ,37 (link)]. The bin width was set to 12 and 0.25 for CT and FDG-PET images, respectively.