Table 2.

Ablation study for MSU-Net and its variants.

Architecture	BUL	EM	NS
	M ± SD	M ± SD	M ± SD
MSU-Net	0.708 ± 0.011	0.900 ± 0.001	0.702 ± 0.010
MSU-Net(37+sum)	0.694 ± 0.020	0.894 ± 0.013	0.683 ± 0.017
MSU-Net(encoder)	0.646 ± 0.061	0.889 ± 0.013	0.679 ± 0.021
MSU-Net(decoder)	0.656 ± 0.027	0.883 ± 0.018	0.661 ± 0.024
MSU-Net(37+concatenated)	0.642 ± 0.036	0.899 ± 0.004	0.674 ± 0.024
MSU-Net(73+concatenated)	0.707 ± 0.061	0.900 ± 0.001	0.667 ± 0.022
MSU-Net(37+dilated)	0.640 ± 0.033	0.877 ± 0.005	0.662 ± 0.013

MSU-Net is MSU-Net (37) in Table 1. MSU-Net (37+ sum) is an MSU-Net with feature fusion by adding. MSU-Net (encoder) and MSU-Net (decoder) are obtained by using multi-scale block to replace the convolution block between encoder and decoder in U-Net. MSU-Net (73+concatenated) and MSU-Net (37+concatenated) are obtained after concatenated the convolution kernel with different receptive fields. MSU-Net (37+dilated) is obtained by dilated convolution. Intersection over Union (IoU) is used as the evaluation metric for comparison. Bold values represent the best results.