The effect of preprocessing filters on predictive performance in radiomics

European Radiology Experimental

Table 4 Areas-under-the-curve at receiver operating characteristics of all the models

Dataset	AUC-ROC Original	AUC-ROC All	AUC-ROC Tuned	Δ_{All—Original}	p_{All—Original}	Δ_{Tuned—Original}	p_{Tuned—Original}	Δ_Tuned—All	p_Tuned—All	Best filter
CRLM	0.7 (0.58–0.82)	0.6 (0.47–0.73)	0.7 (0.58–0.82)	-0.1	0.237	0	1	0.1	0.237	Original
Desmoid	0.72 (0.64–0.79)	0.8 (0.73–0.87)	0.82 (0.76–0.89)	0.08	0.024	0.1	0.006	0.02	0.261	Wavelet
GIST	0.7 (0.63–0.77)	0.77 (0.71–0.83)	0.77 (0.71–0.83)	0.07	0.031	0.07	0.031	0	1	All
HN	0.85 (0.79–0.92)	0.81 (0.74–0.89)	0.86 (0.8–0.92)	-0.04	0.216	0.01	0.651	0.05	0.086	Square root
Lipo	0.74 (0.65–0.83)	0.82 (0.74–0.9)	0.82 (0.74–0.9)	0.08	0.04	0.08	0.04	0	1	All
Liver	0.73 (0.65–0.8)	0.69 (0.61–0.77)	0.75 (0.67–0.82)	-0.04	0.393	0.02	0.573	0.06	0.188	Square root
Melanoma	0.57 (0.45–0.68)	0.53 (0.41–0.64)	0.63 (0.52–0.74)	-0.04	0.179	0.06	0.322	0.1	0.023	Exponential

The best-performing model in terms of area-under-the-curve at receiver operating characteristics (AUC-ROC) is denoted by bold face. Δ denotes the difference between the models; for example, a positive Δ_{All—Original} denotes an improvement using all features over the original feature set. The reported p values correspond to a DeLong test between the denoted models. Statistically significance (p < 0.05) is also denoted by bold face