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

Table 3 Comparison of ablation experiments

	Methods	\(\:{\mathcal{L}}_{seg}\)	\(\:{\mathcal{L}}_{edge}\)	\(\:{\mathcal{L}_{consis}^{data}}\)	\(\:{\mathcal{L}_{consis}^{task}}\)	\(\:{\mathcal{L}_{consis}^{feat}}\)	Dice(%)
I	Seg	√					89.14
II	Seg+Edge	√	√				89.64 (+0.50)
III	Seg+Edge+Data	√	√	√			89.90 (+0.76)
IV	Seg+Edge+Data+Task	√	√	√	√		90.20 (+1.06)
V	Seg+Edge+Data+Feat	√	√	√		√	90.34 (+1.20)
VI	Seg+Edge+Data+Task+Feat (Ours Method)	√	√	√	√	√	90.70^a (+1.56)

^adenotes that our method (VI, best value) is significantly better than the baseline method (I) based on a paired t-test (p-value < 0.05)

ISSN: 1471-2261