Table 5.

Bolus localization performance by consistency. Best results for each metric are highlighted in bold. For the RMSE we reported the median and the inter-quartile range (in parentheses).

	CNN backbone	ry	RMSE
Thin boluses	VGG16	0.340	0.806 (0.516–1.111)
	InceptionV3	0.744	0.276 (0.123–0.471)
	ResNet50	0.945	0.331 (0.204–0.491)
	CNN3	0.699	0.256 (0.116–0.529)
	CNN4	0.787	0.360 (0.235–0.623)
Friedman’s test			X2(4) = 83.12, p < .001
Slightly thick boluses	VGG16	0.199	0.569 (0.391–0.935)
	InceptionV3	0.845	0.332 (0.199–0.499)
	ResNet50	0.915	0.432 (0.370–0627)
	CNN3	0.725	0.273 (0.160–0.654)
	CNN4	0.697	0.323 (0.206–0.575)
Friedman’s test			X2(4) = 84.59, p < .001
Mildly thick boluses	VGG16	0.823	0.574 (0.329–0.811)
	InceptionV3	0.861	0.251 (0.155–0.359)
	ResNet50	0.948	0.275 (0.157–0.348)
	CNN3	0.964	0.186 (0.095–0.265)
	CNN4	0.964	0.221 (0.100–0.341)
Friedman’s test			X2(4) = 139.85, p < .001
Moderately thick boluses	VGG16	0.815	0.655 (0.440–0.985)
	InceptionV3	0.921	0.264 (0.147–0.474)
	ResNet50	0.939	0.204 (0.128–0.280)
	CNN3	0.899	0.158 (0.075–0.276)
	CNN4	0.960	0.180 (0.102–0.295)
Friedman’s test			X2(4) = 170.84, p < .001
Extremely thick boluses	VGG16	0.877	0.513 (0.317–0.755)
	InceptionV3	0.979	0.192 (0.119–0.342)
	ResNet50	0.963	0.337 (0.244–0.490)
	CNN3	0.940	0.132 (0.081–0.212)
	CNN4	0.980	0.137 (0.073–0.264)
Friedman’s test			X2(4) = 178.22, p < .001