AD dataset: OASIS-3

We use a longitudinal 3.0T MRI dataset of 176×240/176×256 T1 brain axial MRI slices containing both normal aging subjects/AD patients, extracted from the open access series of imaging studies-3 (OASIS-3) [55]. The 176×240 slices are zero-padded to reach 176×256 pixels. Relying on clinical dementia rating (CDR) [56], common clinical scale for the staging of dementia, the subjects are comprised of:

Unchanged CDR = 0: Cognitively healthy population;

CDR = 0.5: Very mild dementia (∼ MCI);

CDR = 1: Mild dementia;

CDR = 2: Moderate dementia.

Since our dataset is longitudinal and the same subject’s CDRs may vary (e.g., CDR = 0 to CDR = 0.5), we only use scans with unchanged CDR = 0 to assure certainly healthy scans. As CDRs are not always assessed simultaneously with the MRI acquisition, we label MRI scans with CDRs at the closest date. We only select brain MRI slices including hippocampus/amygdala/ventricles among whole 256 axial slices per scan to avoid over-fitting from AD-irrelevant information; the atrophy of the hippocampus/amygdala/cerebral cortex, and enlarged ventricles are strongly associated with AD, and thus they mainly affect the AD classification performance of machine learning [57]. Moreover, we discard low-quality MRI slices. The remaining dataset is divided as follows:

Training set: Unchanged CDR = 0 (408 subjects/1133 scans/57,834 slices);

Test set: Unchanged CDR = 0 (168 subjects/473 scans/24,278 slices),

CDR = 0.5 (152 subjects/253 scans/13,813 slices),

CDR = 1 (90 subjects/135 scans/7532 slices),

CDR = 2 (6 subjects/10 scans/500 slices).

The same subject’s scans are included in the same dataset. The datasets are strongly biased towards healthy scans similar to MRI inspection in the clinical routine. During training for reconstruction, we only use the training set—structural MRI alone—containing healthy slices to conduct unsupervised learning. We do not use a validation set as our unsupervised diagnosis step is non-trainable.