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

Table 1 Quantitative evaluation of the left atrium segmentation

Method	Scans used		Dice(%)	Jaccard(%)	95HD(voxel)	ASD(voxel)
Method	Labeled	Unlabeled	Dice(%)	Jaccard(%)	95HD(voxel)	ASD(voxel)
V-Net^a	8	0	78.57	66.96	21.20	6.07
V-Net^a	16	0	86.03	76.06	14.26	3.51
V-Net^a	80	0	91.14	83.82	5.75	1.52
UA-MT [38]	8(10%)	72	84.25	73.48	13.84	3.36
SASSNet [55]	8(10%)	72	87.32^b	77.72^b	9.62^b	2.55^b
LG-ER-MT [56]	8(10%)	72	85.54^b	75.12^b	13.29^b	3.77^b
DTC [36]	8(10%)	72	87.51	78.17	8.23	2.36
Ours	8(10%)	72	88.34	79.30	7.92	2.02
UA-MT [38]	16(20%)	64	88.88^b	80.21^b	7.32	2.26^b
SASSNet [55]	16(20%)	64	89.54^b	81.24^b	8.24	2.20^b
LG-ER-MT [56]	16(20%)	64	89.62^b	81.31^b	7.16	2.06^b
DTC [36]	16(20%)	64	89.42^b	80.98^b	7.32^b	2.10^b
Ours	16(20%)	64	90.70	83.09	6.41	1.72

^aindicates the segmentation performance trained with only the labeled data sourced from UA-MT [38]
^bdenotes that our method (best value) is significantly better than the reference method (p-value < 0.05) based on a paired t-test). Since the UA-MT method only provided its performance under the 10% labeled data setting and did not release the model files, a paired t-test comparison with this method was not conducted

ISSN: 1471-2261