Compound W-Net with Fully Accumulative Residual Connections for Liver Segmentation Using CT Images

Table 2

The accuracies of training, validation, testing using normal images, and testing using augmented images for all four types of models (U-Net, original, modified, and compound) using both versions with (32➔ 512) and (64➔ 1024) filters structure. The accuracy represents the value of Dice similarity coefficient.

Model	Filters	Training	Validation	Testing
		Dice	Dice	Normal data Dice	Augmented data Dice
U-Net [16]	32-512	0.9650	0.9800	0.4649	0.6007
Bridge U-Net [17]	32-512	0.9755	0.7065	0.7321	0.6031
Our modified bridge	32-512	0.9785	0.6702	0.7912	0.6010
Our compound model	32-512	0.9752	0.9011	0.8988	0.9442
U-Net [16]	64-1024	0.9740	0.9150	0.5873	0.7593
Bridge U-Net [17]	64-1024	0.9738	0.8775	0.7989	0.7534
Our modified bridge	64-1024	0.9812	0.9250	0.8137	0.8368
Our compound model	64-1024	0.9812	0.9113	0.8303	0.7836