. 2020 Mar;6(1):23–33. doi: 10.18383/j.tom.2019.00021

Table 2.

Comparison of the Networks Trained for Sarcoma Segmentation

Cost	Data	Network	Threshold	Precision	Recall	AUC	Dice	VOE
Dice	Multi-contrast	No skip	0.900	0.833	0.820	0.957	0.827	0.994
	Multi-contrast	Skip	0.995	0.891	0.826	0.979	0.857	0.995
	T2 only	No skip	0.900	0.849	0.787	0.950	0.817	0.994
	T2 only	Skip	0.998	0.906	0.776	0.977	0.836	0.994
Cross Entropy	Multi-contrast	No skip	0.656	0.814	0.858	0.996	0.835	0.994
	Multi-contrast	Skip	0.540	0.869	0.856	0.998	0.863	0.995
	T2 only	No skip	0.636	0.833	0.803	0.992	0.818	0.993
	T2 only	Skip	0.516	0.873	0.849	0.997	0.861	0.995

Values have been calculated from the test set which contains 10% (ie, 7) image sets. In this data set, the network trained with cross entropy loss, skip connections, and on multicontrast images performed best according to 4 of the 5 metrics used in the current study.