Automated cervical cell segmentation using deep ensemble learning

BMC Medical Imaging

Table 2 Performance comparison of base models trained from scratch

Task Type	Model	Dice	Sensitivity	Specificity
Cytoplasm	Unet_resnet34	0.926	0.9275	0.9777
	Unet_densenet121	0.9259	0.9202	0.9801
	UnetPlusPlus_resnet34	0.9229	0.926	0.9762
	UnetPlusPlus_densenet121	0.9289	0.953	0.9708
	DeepLabV3_resnet34	0.9146	0.918	0.9736
	DeepLabV3_resnet50	0.9159	0.9393	0.967
	DeepLabV3Plus_resnet34	0.9282	0.948	0.9721
	DeepLabV3Plus_resnet50	0.924	0.9323	0.9747
	Transunet	/	/	/
	Segformer	0.8717	0.8894	0.9554
Nucleus	Unet_resnet34	0.6299	0.8504	0.9931
	Unet_densenet121	0.6676	0.9017	0.9935
	UnetPlusPlus_resnet34	0.6966	0.9212	0.9941
	UnetPlusPlus_densenet121	0.697	0.9287	0.9941
	DeepLabV3_resnet34	0.5326	0.8715	0.9887
	DeepLabV3_resnet50	0.531	0.8917	0.9881
	DeepLabV3Plus_resnet34	0.5237	0.8127	0.9896
	DeepLabV3Plus_resnet50	0.5593	0.8129	0.9912
	Transunet	/	/	/
	Segformer	/	/	/

In the first column, cytoplasm and nucleus stand for the cytoplasm segmentation task and the nucleus segmentation task, respectively. The symbol “/” indicates that the model is collapsed as it predicts all pixels as negative or positive. Bold values represent the best results

ISSN: 1471-2342