Semi-supervised segmentation of cardiac chambers from LGE-CMR using feature consistency awareness

Table 2 Quantitative evaluation of the left ventricular segmentation(mean±variance)

Method	Scans used		Dice(%)	Jaccard(%)	95HD(voxel)	ASD(voxel)
Method	Labeled	Unlabeled	Dice(%)	Jaccard(%)	95HD(voxel)	ASD(voxel)
V-Net^a	8	0	83.67±9.09	72.86±12.12	7.08±8.65	1.91±2.06
V-Net^a	16	0	87.77±5.93	78.67±8.77	3.07±4.77	0.90±1.27
V-Net^a	80	0	91.38±4.00	84.36±6.33	1.49±0.66	0.37±0.38
UA-MT	8(10%)	72	84.68±7.41^b	74.3±10.47^b	5.99±3.4^b	1.71±1.04^b
DTC	8(10%)	72	85.62±7.41^b	75.52±10.47^b	3.08±3.4^b	0.84±1.04^b
Ours	8(10%)	72	87.22±7.00	77.95±9.97	2.27±1.75	0.61±0.5
UA-MT	16(20%)	64	87.94±5.12^b	78.93±7.73^b	3.30±0.91^b	0.92±0.33^b
DTC	16(20%)	64	88.58±5.12	79.85±7.73	1.88±0.91	0.52±0.33
Ours	16(20%)	64	88.99±4.60	80.45±7.10	1.87±0.84	0.51±0.31

^aindicates the segmentation performance trained with only the labeled data
^bdenotes that our method (best value) is significantly better than the reference method (p-value < 0.05) based on a paired t-test

ISSN: 1471-2261