Table 4.

Coefficients of the linear support vector regression in the ensemble model proposed to discriminate early nasopharyngeal carcinoma (NPC) from benign hyperplasia on MRI. Features were ordered based on their nomination frequency across the five models, as well as their values in the first fold’s support vector regression (SVR1). These features were selected from a pool of 422 features using the proposed boosted-bagged repeated elastic net technique (BB-RENT). The ensemble weights were determined using the reciprocal of the Youden index threshold divided by the number of models in the ensemble, after which the threshold for nasopharyngeal carcinoma was normalized to ≥1.

Imaging Filter	Feature Type	Feature Name	SVR 1	SVR 2	SVR 3	SVR 4	SVR 5
original	shape	SurfaceVolumeRatio	−0.34350	−0.41914	−0.35446	−0.41196	−0.35104
lbp-3D-k	glrlm	LongRunHighGrayLevelEmphasis	−0.23201	−0.23496	−0.19496	−0.23516	−0.18680
original	shape	SurfaceArea	−0.15209	−0.12242	−0.13933	−0.05791	−0.10701
lbp-3D-m2	first-order	Kurtosis	−0.07727	−0.16856	−0.15919	−0.17252	-
log-sigma−0-4492-mm-3D	first-order	Mean	0.08879	-	0.05899	−0.06722	0.15951
original	shape	LeastAxisLength	0.10404	0.11893	0.10616	-	-
exponential	glcm	SumEntropy	0.04001	-	-	−0.00654	0.13320
lbp-3D-m1	first-order	Kurtosis	−0.08219	-	-	-	−0.14464
gradient	first-order	Energy	0.05771	0.01792	-	-	-
lbp-2D	glcm	DifferenceVariance	−0.02907	-	-	-	-
exponential	first-order	Energy	-	−0.05008	-	-	0.02571
exponential	first-order	Variance	-	-	-	-	−0.12470
exponential	glrlm	RunVariance	-	−0.04025	-	-	-
lbp-3D-m1	glcm	ClusterShade	-	-	−0.06451	-	-
lbp-3D-m2	glrlm	ShortRunHighGrayLevelEmphasis	-	0.00498	-	-	-
log-sigma−0-4492-mm-3D	first-order	RobustMeanAbsoluteDeviation	-	-	-	−0.10530	-
original	glrlm	RunEntropy	-	-	-	-	0.05171
Intercepts			0.48176	0.49719	0.49814	0.47659	0.47752
$\mathrm{Ensemble} \mathrm{weight}={\left(threshol{d}_i\times 5\right)}^{-1}$			0.33508	0.36677	0.62281	0.42553	0.33341

SVR = support vector regression, threshold_i = Youden index threshold obtained for each fold.