Table 2.

Performance comparisons and evaluation of different segmentation models

Model	Backbone	Optimizer	Learning rate	Weight decay	Recall	Precision	Mean intersection over union	Mean pixel accuracy
CABM-HRNet	hrnetv2_w32	Adam	0.0005	0	0.9116 ± 0.099	0.9204 ± 0.107	0.8521 ± 0.034	0.9116 ± 0.099
	hrnetv2_w18	Adam	0.0005	0	0.9061 ± 0.097	0.9174 ± 0.138	0.8500 ± 0.027	0.9061 ± 0.097
	hrnetv2_w32	SGD	0.004	0.0001	0.8953 ± 0.099	0.9122 ± 0.123	0.8360 ± 0.023	0.8953 ± 0.099
	hrnetv2_w18	SGD	0.004	0.0001	0.8939 ± 0.100	0.9053 ± 0.168	0.8302 ± 0.029	0.8939 ± 0.100
HRNet	hrnetv2_w32	Adam	0.0005	0	0.9100 ± 0.110	0.9189 ± 0.131	0.8510 ± 0.042	0.9100 ± 0.110
	hrnetv2_w18	Adam	0.0005	0	0.9097 ± 0.125	0.9176 ± 0.142	0.8505 ± 0.040	0.9097 ± 0.125
	hrnetv2_w32	SGD	0.004	0.0001	0.8910 ± 0.126	0.9140 ± 0.130	0.8341 ± 0.043	0.8910 ± 0.126
	hrnetv2_w18	SGD	0.004	0.0001	0.8896 ± 0.175	0.9064 ± 0.139	0.8278 ± 0.056	0.8896 ± 0.175
PSPNet	MobileNetv2	Adam	0.0005	0	0.8995 ± 0.149	0.8883 ± 0.124	0.8221 ± 0.052	0.8995 ± 0.149
	ResNet50	Adam	0.0005	0	0.9018 ± 0.230	0.8939 ± 0.119	0.8278 ± 0.052	0.9018 ± 0.230
	MobileNetv2	SGD	0.01	0.0001	0.8566 ± 0.498	0.8554 ± 0.177	0.7718 ± 0.239	0.8566 ± 0.498
	ResNet50	SGD	0.01	0.0001	0.8777 ± 0.315	0.8900 ± 0.148	0.8082 ± 0.122	0.8777 ± 0.315
DeeplabV3 +	MobileNetv2	Adam	0.0005	0	0.9101 ± 0.151	0.9118 ± 0.147	0.8468 ± 0.051	0.9101 ± 0.151
	Xception	Adam	0.0005	0	0.9060 ± 0.135	0.9100 ± 0.143	0.8425 ± 0.038	0.9060 ± 0.135
	MobileNetv2	SGD	0.007	0.0001	0.8994 ± 0.265	0.8826 ± 0.148	0.8178 ± 0.064	0.8994 ± 0.265
	Xception	SGD	0.007	0.0001	0.8945 ± 0.265	0.8843 ± 0.109	0.8158 ± 0.055	0.8945 ± 0.265
U-Net	ResNet50	Adam	0.0001	0	0.9055 ± 0.103	0.9172 ± 0.145	0.8473 ± 0.035	0.9055 ± 0.103
	VGG	Adam	0.0001	0	0.9045 ± 0.044	0.9198 ± 0.162	0.8484 ± 0.014	0.9045 ± 0.044
	ResNet50	SGD	0.01	0.0001	0.8892 ± 0.303	0.8944 ± 0.142	0.8192 ± 0.084	0.8892 ± 0.303
	VGG	SGD	0.01	0.0001	0.8948 ± 0.084	0.9057 ± 0.137	0.8312 ± 0.020	0.8948 ± 0.084

Different backbone, optimizers, and learning rates were used according to different segmentation models. The evaluation indicators were measured on the test set with ten-fold cross-validation (mean ± standard deviation). The best results of each network are shown in bold