**3. Results**

Patient characteristics of included patients with central necrosis (*n* = 43) are presented in Table 1. In the NSCLC cohort, central necrosis was observed in 12 out of 35 patients (34%). In the PPGL cohort, central necrosis was observed in 31 of 77 patients (40%; Figure 1).

**Table 1.** Clinical characteristics of 31 PPGL and 12 NSCLC patients with central necrosis. SUV: standardised uptake value, MTV: metabolic tumour volume, NTF: necrotic tumour fraction, NSCLC: non-small cell lung carcinomas, PPGL: pheochromocytomas and paragangliomas.


**Figure 1.** Example of a [18F]FDG PET/CT scan of a patient with a pheochromocytoma with central necrosis in the left adrenal gland, with VOIvital-tumour in blue and VOIgross-tumour in yellow. The NTF is 0.05.

At least 65% of the features were affected by the choice of delineation (Table 2). The PPGL population was influenced the most with 79% of affected features, compared to 65% in the NSCLC population, which is a significant difference between the populations (*p* = 0.031). Out of the 105 features, 61% were affected in the NSCLC cohort as well as the PPGL cohort. For all patients taken together, even 82% of the features were affected.

First order features were affected substantially, with at least 72% of features, followed by texture features with at least 66% of features. Shape features were affected the least, with 50% of affected features in both datasets. Of all texture feature classes, GLCM features were affected the least, with a maximum of 68% of features. For all other classes, the maximum number of affected features was at least 80%.

**Table 2.** Numbers and percentages of features affected by the delineation method (VOIvital-tumour vs. VOIgross-tumour) per feature class for the different cohorts and both cohorts together. Differences in the number of affected features between cohorts were assessed using the Fisher's exact test. NSCLC: non-small cell lung carcinoma, PPGL: pheochromocytoma and paraganglioma, GLCM: grey level cooccurrence matrix, GLRLM: grey level run length matrix, GLSZM: grey level size zone matrix, GLDM: grey level dependence matrix, NGTDM: neighbouring grey tone difference matrix.


The size of the NTF appeared to influence the number of affected shape features (nonsignificant; Table 3). For a small NTF, 36% of the features were affected, increasing to 57% and 71% for medium and large NTFs, respectively. Moreover, for small and medium NTFs 100% of the NGTDM features were affected compared to only 40% for a large NTF.

**Table 3.** Numbers and percentages of features affected by the delineation method (VOIvital-tumour vs. VOIgross-tumour) per feature class for the subgroups based on NTF. Differences in the number of affected features between subgroups were assessed using the Fisher's exact test. NTF: necrotic tumour fraction, GLCM: grey level cooccurrence matrix, GLRLM: grey level run length matrix, GLSZM: grey level size zone matrix, GLDM: grey level dependence matrix, NGTDM: neighbouring grey tone difference matrix.


Although nonsignificant, the value of the SUVmax appeared to influence the number of affected features (Table 4) and a higher value resulted in more affected features (56%, 66% and 70% for low, medium and high values, respectively). This increasing trend could also be observed for all texture feature classes except for GLCM features, where the number of affected features decreased with an increasing SUVmax.

Overlap in affected features between cohorts and subgroups is generally high (Figure 2), ye<sup>t</sup> the affected shape features varied largely between cohorts and for different-sized NTFs. It can be observed that many features that were affected in one subset or cohort, were also affected in most of the other subsets or cohorts (Supplementary Table S1).

**Table 4.** Numbers and percentages of features affected by the delineation method (VOIvital-tumour vs. VOIgross-tumour) per feature class for the different subgroups based on SUVmax. Differences in the number of affected features between subgroups were assessed using the Fisher's exact test. SUVmax: maximum standardised uptake value, GLCM: grey level co-occurrence matrix, GLRLM: grey level run length matrix, GLSZM: grey level size zone matrix, GLDM: grey level dependence matrix, NGTDM: neighbouring grey tone difference matrix.


**Figure 2.** Venn diagrams representing the overlap in affected features per feature class for the cohorts (**a**) and subgroups ((**b**): NTF, (**c**): SUVmax). NSCLC: non-small cell lung carcinoma, PPGL: pheochromocytoma and paraganglioma, NTF: necrotic tumour fraction, SUVmax: maximum standardised uptake value.

For each of the three radiomic models evaluating predictive performance (VOIvital-tumour, VOIgross-tumour and combined) for the PPGL dataset, three factors were retained and the three best corresponding features were selected (Table 5). The KMOs of the models were excellent (>0.96). AUCs varied 0.791–0.829, but were not significantly different between the radiomic models (VOIvital-tumour vs. VOIgross-tumour: *p* = 0.775; VOIvital-tumour vs. combined: *p* = 0.625; VOIgross-tumour vs. combined: *p* = 0.874; Figure 3). The mean AUCs of the sham experiments were lower (0.645–0.655) than the AUCs of the radiomic models, indicating the validity of the findings.

**Table 5.** Results of dimension reduction and predictive performance of the VOIvital-tumour, VOIgross-tumour and combined model for the noradrenergic biochemical profile in the PPGL cohort. For each model, 3 features were selected corresponding to the factors with the highest loadings. AUCs of the radiomic models and the mean AUC of the sham experiment are reported. Features marked with \* are not significantly different between delineation methods in the PPGL dataset. KMO: Kaiser–Meier–Olkin (KMO) measure, AUC: area under the receiver operating characteristic curve.


**Figure 3.** ROC curves and AUCs for the different radiomic models: blue: VOIvital-tumour (features: first order minimum, shape surface area and GLCM informational measure of correlation 2); green: VOIgross-tumour (features: shape surface area, NGTDM complexity and GLDM dependence entropy); black: combined (features: VOIgross-tumour GLCM sum entropy, VOIvital-tumour shape maximum 3D diameter and VOIvital-tumour shape surface volume ratio). ROC: receiver operating characteristic, AUC: area under the ROC curve.
