Table 3. Comparison with SOTA 2D and 3D methods on RETOUCH Cirrus and Spectralis dataset with the 3-fold cross-validation method.

Network	Scanner				Complexity
	Cirrus		Spectralis
Full-supervised	DSC(%)↑	Jaccard(%)↑	DSC(%)↑	Jaccard(%)↑	Parameters
UNet [10]	73.24 ± 7.65	63.38 ± 6.03	72.75 ± 8.91	63.25 ± 6.57	5.71 M
UNet++[11]	77.51 ± 6.01	67.04 ± 5.80	79.54 ± 5.46	68.14 ± 5.73	9.04 M
MsTGANet [18]	78.84 ± 5.59	67.31 ± 5.72	81.32 ± 4.17	70.29 ± 4.81	11.61 M
RetifluidNet [19]	83.27 ± 4.63	73.07 ± 3.43	83.16 ± 3.85	72.88 ± 3.60	34.27 M
UNet(3D) [33]	84.87 ± 4.75	73.70 ± 3.22	82.84 ± 4.13	72.59 ± 3.86	12.94 M
ResNet34(3D) [34]	87.32 ± 2.74	74.69 ± 2.87	85.04 ± 3.76	73.80 ± 3.14	17.93 M
Semi-supervised (23%)
UAMT [22]	83.15 ± 4.83	72.07 ± 3.81	77.68 ± 6.20	64.33 ± 5.97	9.44 M
URPC [36]	83.76 ± 4.09	72.61 ± 3.56	77.15 ± 6.31	63.42 ± 6.04	5.89 M
MCNet [35]	87.24 ± 2.81	77.70 ± 2.33	85.33 ± 3.52	74.16 ± 3.03	12.35 M
CAML [37]	85.34 ± 3.67	76.02 ± 2.54	81.04 ± 4.68	72.09 ± 4.18	20.13 M
Ours	89.83 ± 2.04	81.36 ± 1.41	88.54 ± 2.73	79.57 ± 2.45	15.64 M