Table 3.

Comparison of U-Net and its variants and the proposed Half-UNet on three datasets.

Architecture	Params	FLOPs	Mammography Dice	Lung nodule Dice	Endocardium Dice	Epicardium Dice
U-Net	31.04 M	11×	0.8939	0.8842	0.8797	0.9299
UNet3+	26.97 M	43×	0.8920	0.8864	0.8633	0.9316
DC-UNet	10.07 M	6×	0.8940	0.8855	0.9059	0.9503
Half-UNet*†_u	20.03 M	20×	0.8911	0.8873	0.8691	0.8976
Half-UNet*†_d	38.09 M	7×	0.8922	0.8853	0.8926	0.9107
Half-UNet†	0.41 M	2×	0.8944	0.8858	0.8794	0.9281
Half-UNet	0.21 M	1×	0.8892	0.8821	0.9122	0.9555
			Sensitivity	Sensitivity	Sensitivity	Sensitivity
U-Net	31.04 M	11×	0.8745	0.9037	0.8475	0.9097
UNet3+	26.97 M	43×	0.8738	0.9033	0.8345	0.9134
DC-UNet	10.07 M	6×	0.8804	0.9046	0.8906	0.9310
Half-UNet*†_u	20.03 M	20×	0.8725	0.8971	0.8547	0.8877
Half-UNet*†_d	38.09 M	7×	0.8763	0.8916	0.8914	0.9027
Half-UNet†	0.41 M	2×	0.8875	0.9131	0.8773	0.9209
Half-UNet	0.21 M	1×	0.8821	0.9208	0.9029	0.9488
			Specificity	Specificity	Specificity	Specificity
U-Net	31.04 M	11×	0.9942	0.9941	0.9995	0.9991
UNet3+	26.97 M	43×	0.9939	0.9939	0.9995	0.9991
DC-UNet	10.07 M	6×	0.9934	0.9945	0.9995	0.9994
Half-UNet*†_u	20.03 M	20×	0.9938	0.9946	0.9994	0.9989
Half-UNet*†_d	38.09 M	7×	0.9933	0.9949	0.9993	0.9989
Half-UNet†	0.41 M	2×	0.9926	0.9931	0.9992	0.9989
Half-UNet	0.21 M	1×	0.9923	0.9925	0.9994	0.9990

The symbol ^† means that the Ghost module is not used, and * indicates that the numbers of channels are not unified. The best results are highlighted in bold. “_u” represents feature fusion using the Upsampling2D + 3 × 3 convolution strategy. “_d” represents feature fusion using the deconvolution strategy.