Table 3.

Pharyngeal phase detection performance by bolus consistency. The best results for each metric are highlighted in bold. P3_BPM and P3_UESC values higher than or equal to the inter-rater agreement are marked with an asterisk.

	CNN backbone	F1-score	P3BPM (%)	P3UESC (%)
Thin boluses	VGG16	0.921	89.66	89.66*
	InceptionV3	0.927	89.66	93.10*
	ResNet50	0.934	89.66	89.66*
	CNN3	0.893	89.66	75.86
	CNN4	0.901	89.66	82.76
Inter-rater agreement			93.10	86.21
Slightly thick boluses	VGG16	0.914	92.54	94.09
	InceptionV3	0.904	89.55	92.54
	ResNet50	0.927	89.55	100.00*
	CNN3	0.901	86.57	86.57
	CNN4	0.894	88.06	91.04
Inter-rater agreement			94.06	95.52
Mildly thick boluses	VGG16	0.881	75.36	91.30*
	InceptionV3	0.850	75.36	88.41*
	ResNet50	0.869	76.81	89.86*
	CNN3	0.848	71.01	85.51
	CNN4	0.868	75.36	89.86*
Inter-rater agreement			86.96	88.41
Moderately thick boluses	VGG16	0.925	77.94	100.00
	InceptionV3	0.913	72.06	97.06*
	ResNet50	0.925	82.35*	97.06*
	CNN3	0.875	70.59	88.24*
	CNN4	0.890	69.12	94.12*
Inter-rater agreement			82.35	88.24
Extremely thick boluses	VGG16	0.900	64.41	96.61*
	InceptionV3	0.860	61.02	91.53*
	ResNet50	0.908	69.49	93.22*
	CNN3	0.830	64.41	84.75
	CNN4	0.876	61.02	94.92*
Inter-rater agreement			81.36	88.14