Table 2:

Ablation analysis of our method over loss layers, projection+prediction layer widths (#MLP), loss functions $\left({ℒ}_{Rec},{ℒ}_{cs}\right)$, use of augmentation, and hyperparameters $\left(\beta ,\mu ,\gamma \right)$.

Exp	Loss Layers	#MLP	${ℒ}_{Rec}$	Aug.	$\beta$	$\mu$	$\gamma$	${ℒ}_{cs}$	IBIS-subcort	IBIS-wmgm	OASIS3
A	Enc	256							0.829(0.068)	0.733(0.062)	0.783(0.16)
B	Enc	2048							0.849(0.060)	0.732(0.073)	0.809(0.13)
C	Enc	2048	✓						0.859(0.058)	0.713(0.079)	0.811(0.13)
D	EncDec	2048	✓						0.860(0.058)	0.718(0.066)	0.810(0.13)
E	EncDec	2048		✓					0.858(0.060)	0.724(0.077)	0.809(0.13)
F	EncDec	2048	✓	✓					0.856(0.060)	0.739(0.067)	0.812(0.13)
G	EncDec	2048	✓		100				0.857(0.062)	0.728(0.074)	0.809(0.13)
H	EncDec	2048	✓			10−3	10−3		0.845(0.063)	0.739(0.074)	0.804(0.13)
I	EncDec	2048	✓		100	10−3	10−3		0.859(0.056)	0.735(0.061)	0.811(0.13)
J	EncDec	2048	✓	✓	100	10−3	10−3		0.863(0.057)	0.758(0.062)	0.808(0.14)
K	EncDec	2048	✓	✓	100	10−2	10−3		0.853(0.055)	0.745(0.058)	0.813(0.13)
L	EncDec	2048	✓	✓	100	*	10−3	✓	0.870(0.052)	0.806(0.030)	0.810(0.13)

Mean dice is used for quantification on all datasets.

^*$\mu ={10}^{-3}$ on IBIS- {wmgm, subcort} and $\mu ={10}^{-2}$ on OASIS3.