Table 1.
Scanning parameters for all scans included in the study.
Table 2.
Patient characteristics for all four datasets used in this study. The bold numbers 1, 2, 3 or 4 indicate the datasets from which the variable was significantly different.
Fig 1.
Kaplan-Meier curves for overall survival of all datasets.
Fig 2.
Receiver Operator Curves (ROC) for the cohort difference (CD) model for each combination of datasets, including three radiomic features and ‘two-year survival’ as independent variables.
Fig 3.
Prognostic index (PI) ranges for all datasets based on the model developed on CT–scan1, using Dataset 1 as training dataset.
Table 3.
Values of Harrell’s concordance index with 95% confidence intervals, in the case Dataset 1 (n = 50) was used as training (T) to develop a model using LASSO. Validation (V) results are shown for Dataset 2, 3 and 4. Significant values are indicated in grey. A hyphen indicates that either all coefficients were forced to zero, or all predictions were equal to one, meaning that no linear combination of any subset of regressors was useful in predicting the outcomes.
Fig 4.
Heatmap with the mean AUC values found for each classifier and each image set, predicting two-year survival. The entire cohort with at least two-years of follow-up (n = 134) was used in this investigation.
Table 4.
Cox regression on the percentage variation of the PET imaging descriptors most commonly used, reporting the univariable hazard ratio (HR), 95% confidence interval (CI) of the HR and corresponding p-value. Univariable performance is reported in terms of the concordance-index (c-index). Absolute values of Scan 1 and 2, and percentage variation between PET acquisitions of the analyzed metrics are also presented (mean ± standard deviation).