Table VII.

Evaluation of the segmentation results on ACDC dataset for RV blood-pool, LV myocardium, and LV blood-pool (mean values reported), obtained from different weighting schemes of the categorical cross-entropy loss function.

	End diastole (ED)					End systole (ES)
	None	Class	Spatial	Spatial w/Class	DMR UNet	None	Class	Spatial	Spatial w/Class	DMR UNet
Dice (%)	91.6	89.2	91.7	91.8	92.2	87.2	84.7	88.1	87.8	88.8
Jaccard (%)	85.0	81.2	85.1	85.2	85.9	78.3	74.6	79.3	79.0	80.4
MSD (mm)	0.54	0.71	0.53	0.51	0.53	1.08	1.25	0.89	0.95	0.83
HD (mm)	10.03	10.48	10.06	9.99	9.52	12.61	12.60	11.31	12.16	10.96

UNet model trained with cross-entropy loss: without any weighting, class frequency weighting, spatial weighting (with uniform class weight), and spatial with class frequency weighting, compared against the proposed DMR-UNet model.

Best performing model for the ED and ES phases are shown in bold case.