TABLE 3.

Experimental scenarios in looser datasets.

Scenario	Train objectives		Test objectives		Test accuracy (%)
	Data component	Data amount	Data component	Data amount
1	FTD, FTD_NC	440, 282	FTD, FTD_NC	112, 72	93.45
2	AD, AD_NC	334, 381	AD, AD_NC	88, 88	89.86
3	FTD, AD, NC	440, 334, 282+382	FTD, AD, NC	112, 88, 72+88	91.83
4	FTD, AD, NC	440, 334, 282+382	FTD, AD	112, 88	93.05
5	FTDl, FTD_NCl	1,001, 475	FTDl, FTD_NCl	249, 122	68.02
6	ADl, AD_NCl	1,051, 749	ADl, AD_NCl	263, 189	77.18
7	FTDl, ADl, NCl	1,001, 1,051, 475+749	FTDl, ADl, NCl	249, 263, 122+189	66.79
8	FTDl, ADl, NCl	1,001, 1,051, 475+749	FTDl, ADl	249, 263	81.25
9	FTDl, ADl, NCl	1,001, 1,051, 475+749	FTD, AD	112, 88	98.61

Among them, the train dataset of scenario 7 was the same as that of scenarios 8 and 9. This design was intended to evaluate the generalizability of the network. The test datasets of scenario 9 and scenario 4 overlapped to further measure the robustness of the network. See Table 2 for the sample size of each data component. For example, in scenario 6 of Table 3, the training sample size and the test sample size corresponding to AD^l are shown in the first row in the main body of Table 2, i.e., 1,051 and 253, respectively. The superscript “l” indicates that we removed the restriction of scanning from MPRAGE sequence and distinguish from scenarios 1–4.