2.2.1. Image Processing

For NSCLC, images were interpolated to isotropic voxels of 3.38 × 3.38 × 3.38 mm<sup>3</sup> using trilinear interpolation with the grids aligned by the centre using MATLAB version 2017b (Mathworks, Natick, MA, USA) [14]. PPGL images were not interpolated since the voxels were almost isotropic (3.18 × 3.18 × 3.00 mm3).

#### 2.2.2. Volumes of Interest Delineation

VOIs were delineated semi-automatically using 3DSlicer version 4.11 (www.slicer.org, accessed on 1 March 2021) [20] and in-house built software implemented in Python version 3.7 (Python Software Foundation, Wilmington, Delaware). The often peripheral region of the tumour showing increased [18F]FDG uptake (VOIvital-tumour) was delineated using a semi-automatic threshold-based method and corrected for local background [10]. A threshold of 41% of the peak standardised uptake value (SUVpeak) that was obtained using a sphere of 12 mm diameter [21] was selected, since the delineated tumour sizes of this method agreed best with pathological tumour sizes [22]. As tumours in the PPGL cohort showed low contrast between the tumour and surrounding tissue, boxing was applied to exclude the surrounding [18F]FDG-avid tissues. VOIgross-tumour was generated by manually adding the volumes of central necrosis to VOIvital-tumour, using the low-dose CT as a visual reference. The necrotic tumour fraction (NTF) was determined using Equation (1):

$$\text{NTF} = 1 - \text{VOI}\_{\text{virtual-turnour}} / \text{VOI}\_{\text{gross-turnour}} \tag{1}$$

where a NTF of 0 indicates no central necrosis and a higher NTF indicates larger volumes of necrosis. Patients were selected when NTF > 0.

#### 2.2.3. Radiomic Feature Extraction

For each selected patient, 105 radiomic features were extracted from VOIvital-tumour and VOIgross-tumour using PyRadiomics version 3.0 in Python version 3.7 (Python Software Foundation, Wilmington, DE, USA) [23]: 18 first order features, 14 shape features, 22 grey level cooccurrence matrix (GLCM) features, 16 grey level run length matrix (GLRLM) features, 16 grey level size zone matrix (GLSZM) features, 14 grey level dependence matrix (GLDM) features and 5 neighbouring grey tone difference matrix (NGTDM) features. A fixed bin size of 0.5 g/mL was applied.
