Table A.5.

Comparison of model performance averaged across 10 test sets. Performance was measured with structural similarity index (SSIM) and peak signal-to-noise ratio (PSNR) in decibels (dB) on the generated MRI images with the skull removed (i.e., voxels within the group template brain mask only). Performance was evaluated overall and separately by clinical group (CN, MCI, and AD). P indicates the imputation method conditioned on the most recent past image. PF indicates the imputation method conditioned on both the most recent past image and the closest following image. AE indicates imputation using an autoencoder. Bold values indicate the best performing imputation method for a given clinical diagnosis group within the P or PF condition.

Model	Images	CN	MCI	AD	ALL
SSIM (higher=better)
Naive-P	skull removed	0.899 ± 0.015	0.899 ± 0.005	0.914 ± 0.007	0.901 ± 0.005
AE-P	skull removed	0.913 ± 0.008	0.912 ± 0.006	0.911 ± 0.002	0.912 ± 0.006
ReMiND-P	skull removed	0.974 ± 0.003	0.974 ± 0.001	0.974 ± 0.002	0.974 ± 0.001
Naive-PF	skull removed	0.894 ± 0.025	0.886 ± 0.013	0.893 ± 0.008	0.890 ± 0.006
AE-PF	skull removed	0.919 ± 0.006	0.914 ± 0.0002	0.916 ± 0.005	0.916 ± 0.002
ReMiND-PF	skull removed	0.984 ± 0.002	0.982 ± 0.0004	0.983 ± 0.0009	0.983 ± 0.0002
PSNR (dB) (higher=better)
Naive-P	skull removed	22.868 ± 0.754	24.644 ± 0.713	23.217 ± 0.632	23.217 ± 0.632
AE-P	skull removed	25.911 ± 0.768	26.820 ± 0.314	26.962 ± 0.660	26.962 ± 0.660
ReMiND-P	skull removed	29.078 ± 0.781	28.908 ± 0.458	28.824 ± 0.651	28.960 ± 0.434
Naive-PF	skull removed	23.743 ± 2.957	22.537 ± 1.353	23.457 ± 1.026	23.097 ± 0.571
AE-PF	skull removed	27.040 ± 0.590	26.613 ± 0.100	27.954 ± 0.662	27.792 ± 0.232
ReMiND-PF	skull removed	31.786 ± 0.399	31.175 ± 0.218	31.857 ± 0.343	31.449 ± 0.090