Enhanced segmentation of gastrointestinal polyps from capsule endoscopy images with artifacts using ensemble learning

Table 1 Comparison between the single and ensemble models on GZ_Capcam (artifact-affected) datasets

Model	SegNet		U-Net		Attention-UNet		ResNet-UNet		HarDMSEG
Metric	IoU	Dice	IoU	Dice	IoU	Dice	IoU	Dice	IoU	Dice
Snap_max	0.341	0.454	0.379	0.507	0.361	0.502	0.414	0.522	0.538	0.649
Ens_max	0.341	0.462	0.406	0.532	0.375	0.509	0.435	0.537	0.550	0.653
Improve	0.000	0.008	0.027	0.025	0.014	0.007	0.021	0.015	0.012	0.004
Improve (%)	0.08	1.71	7.01	4.93	3.88	1.40	5.13	2.86	2.22	0.61

Snap_max and Ens_max denote the performance of the best performing single model and ensemble model, respectively, and the last two rows denote the improvement of the ensemble learning model compared to a single model. GZ_Capcam: The test set of experimental images from Guangzhou First People’s Hospital. IoU: Intersection over union.

Table 2 Comparison between the single and ensemble models on CVC_Colon (clear) datasets

Model	SegNet		UNet		Attention-UNet		ResNet-UNet		HarDMSEG
	IoU	Dice	IoU	Dice	IoU	Dice	IoU	Dice	IoU	Dice
Snap_max	0.700	0.780	0.713	0.788	0.754	0.830	0.747	0.819	0.840	0.901
Ens_max	0.702	0.779	0.711	0.783	0.752	0.829	0.750	0.816	0.840	0.901
Improve	0.002	-0.001	-0.002	-0.005	-0.002	-0.002	0.003	-0.002	0.000	0.000
Improve (%)	0.30	-0.08	-0.28	-0.61	-0.22	-0.22	0.42	-0.26	0.03	-0.01

Snap_max and Ens_max denote the performance of the best performing single model and ensemble model, respectively, and the last two rows denote the improvement of the ensemble learning model compared to a single model. IoU: Intersection over union.

Table 3 Comparison between the single and ensemble models on CVC_Clinic (clear) datasets

Model	SegNet		UNet		Attention-UNet		ResNet-UNet		HarDMSEG
	IoU	Dice	IoU	Dice	IoU	Dice	IoU	Dice	IoU	Dice
Snap_max	0.814	0.890	0.816	0.884	0.836	0.900	0.823	0.884	0.845	0.893
Ens_max	0.815	0.890	0.826	0.891	0.838	0.898	0.824	0.884	0.844	0.892
Improve	0.001	0.001	0.010	0.007	0.002	-0.001	0.001	0.000	-0.001	-0.001
Improve (%)	0.13	0.07	1.20	0.76	0.23	-0.14	0.08	0.02	-0.08	-0.06

Snap_max and Ens_max denote the performance of the best performing single model and ensemble model, respectively, and the last two rows denote the improvement of the ensemble learning model compared to a single model. IoU: Intersection over union.