Table 3.

The discrimination performance of models in the training and testing datasets.

Model	Cutoff	Training dataset				Testing dataset				P value
		AUC (95% CI lower–upper)	SPE	SEN	ACC	AUC (95% CI lower–upper)	SPE	SEN	ACC
T2WI score	− 0.768	0.769 (0.639–0.900)	0.659	0.810	0.708	0.758 (0.569–0.946)	0.759	0.800	0.759	0.922
DWI score	− 0.578	0.634 (0.483–0.785)	0.886	0.789	0.723	0.653 (0.440–0.866)	0.789	0.200	0.586	0.890
CT score	− 0.713	0.874 (0.777–0.972)	0.773	0.857	0.800	0.821 (0.650–0.993)	0.947	0.600	0.828	0.598
MR score	− 0.555	0.788 (0.664–0.912)	0.795	0.714	0.769	0.805 (0.635–0.976)	0.789	0.600	0.724	0.872
Integrated score	− 0.656	0.903 (0.828–0.977)	0.864	0.810	0.846	0.889 (0.770–1.000)	0.895	0.500	0.759	0.856
Concise score	− 1.254	0.906 (0.833–0.979)	0.773	0.905	0.815	0.884 (0.761–1.000)	0.895	0.600	0.793	0.768

When the model value lesser than the corresponding cut-off value means perineural invasion status is negative; AUC the area under receiver operating characteristic (ROC) curves, 95% CI lower–upper the lower value to the upper value of 95% confidence interval, SPE specificity, SEN sensitivity, ACC accuracy; P value was derived from Delong test between the training and testing datasets.