Table 4. Ablation study of our proposed GRA module and CML training strategy on the DME dataset.

Training		Designs			Metrics
Labeled	Unlabeled	GRA	CML	DSC(%)↑	Jaccard(%)↑	95HD(voxel)↓	ASD(voxel)↓
4(10%)	36(90%)			75.90 ± 6.13	61.36 ± 6.41	7.81 ± 5.51	2.87 ± 1.89
		√		83.64 ± 3.22	72.45 ± 3.43	4.56 ± 1.57	1.55 ± 1.32
			√	86.36 ± 1.24	75.80 ± 1.69	2.41 ± 0.83	0.93 ± 0.34
		√	√	87.50 ± 0.91	77.79 ± 1.43	2.77 ± 1.35	0.79 ± 0.38
10(25%)	30(75%)			83.13 ± 4.42	70.79 ± 5.29	5.34 ± 3.44	1.54 ± 1.21
		√		85.62 ± 1.53	74.78 ± 1.86	2.67 ± 0.35	0.79 ± 0.37
			√	88.36 ± 1.41	79.29 ± 1.53	1.74 ± 0.32	0.68 ± 0.29
		√	√	91.04 ± 0.24	81.93 ± 0.24	1.38 ± 0.29	0.56 ± 0.27