Table 4.

Classification accuracy for sMCI vs. pMCI (ADNI data).

Methods	sMCI vs. pMCI
	Frozen			Fine-tuned
	BACC	AUC	F1	BACC	AUC	F1
No pretrain	–	–	–	69.3	71.6	70.9
AE	62.6	65.4	62.8	69.5	71.8	71.1
VAE (Kingma and Welling, 2013)	61.3	64.8	62.9	63.8	65.9	64.3
SimCLR (Chen et al., 2020)	63.3	66.3	64.4	69.5	71.9	70.6
MoCo (He et al., 2020)	64.6	66.5	65.7	70.8	72.4	71.4
BYOL (Grill et al., 2020)	64.2	66.4	64.9	70.3	72.2	71.4
LSSL (Zhao et al., 2021a)	69.4	71.8	70.5	71.2	73.7	72.8
LNE	71.1	73.7	71.8	73.5	75.6	74.4

The highest accuracy scores are in bold. The classifier based on the LNE encoding was significantly more accurate than the alternative methods for both frozen and fine-tuned encoder ($p<0.05$, DeLong’s test).