Preselection of robust radiomic features does not improve outcome modelling in non-small cell lung cancer based on clinical routine FDG-PET imaging

EJNMMI Research

Table 2 Results of the multivariable analysis

Outcome	Radiomic features (feature type)	Robustness delineation (ICC)	Robustness attenuation correction (ICC)	Robustness motion (ICC)	AUC training (range)	AUC validation (95% CI)
12-month EFS	LHL coefficient of variation (wavelet)	0.79	0.81	0.00	0.73 (0.65–0.81)	0.74 (0.55–0.93)
	LHH neighbouring grey-level dependence matrix high dependence high grey-level emphasis (wavelet)	0.91	0.35	0.92
	HLH skewness (wavelet)	0.58	0.37	0.43
24-month EFS	HLH mean (wavelet)	0.49	0.17	0	0.74 (0.58–0.94)
	LHH neighbouring grey-level dependence matrix high dependence high grey-level emphasis (wavelet)	0.91	0.35	0.92
12-month OS	LLH skewness (wavelet)	0.74	0.06	0.47	0.85 (0.6–1)	0.67 (0.43–0.91)
	HLL kurtosis (wavelet)	0.93	0.93	0.75
	HHL skewness (wavelet)	0.85	0.00	0.49
	LLH grey-level run length matrix short run high grey-level emphasis (wavelet)	1.00	0.83	0.93
24-month OS	HLL skewness (wavelet)	0.82	0.91	0.39	0.69 (0.57–0.8)
12-month OS robust preselection	HHL NGLDM dependence count entropy (wavelet)	0.95	0.98	0.96	0.67 (0.46–0.85)	0.53 (0.26–0.81)

Results multivariable analysis for EFS and OS. Radiomic features were selected using backward selection. Good classification performances of models without robust preselection were observed for the training cohort (AUC = 0.69–0.85) and the validation cohort (AUC = 0.67–0.74). Performance of the robust model was moderate in training (AUC = 0.67) and weak in validation (AUC = 0.53)