Table 2:

Summary of polyp classification results for the architectures described earlier (see Table 1), obtained by 7-fold cross-validation on the 8641 colonoscopy images. The sensitivity (true positive rate) is given at false negative rates of 5% and 1% (i.e. a specificity of 95% and 99% respectively), as sensitivity and specificity are interdependent values.

Model	Initial Weights	Accuracy	AUC	Sensitivity at 5% FNR	Sensitivity at 1% FNR
NPI-CNN 1	-	91.9 ± 0.2%	0.970 ± 0.002	88.1%	65.4%
NPI-CNN 2	-	91.0 ± 0.4%	0.966 ± 0.002	86.2%	60.6%
PI-CNN 1	VGG16	95.9 ± 0.3%	0.990 ± 0.001	96.9%	87.8%
PI-CNN 2	VGG19	96.4 ± 0.3%	0.991 ± 0.001	96.9%	88.1%
PI-CNN 3	ResNet50	96.1 ± 0.1%	0.990 ± 0.001	96.8%	88.0%