FIGURE 2.

The comprehensive workflow of the generative adversarial network constrained multiple loss autoencoder (GANCMLAE) model. (a) Two cohorts were enrolled. Cohort A was from the ADNI and cohort B were from the SILCODE project. (b) The procedures of structural magnetic resonance imaging (MRI) processing for model training. (c) Train the model with normal controls (NCs) and then reconstruct the input data including NC, Alzheimer's disease (AD), or mild cognitive impairment (MCI). Residual scans can be gained from the input and output. (d) Evaluate the performance of the model from two aspects: a. structural similarity index measure (SSIM) of original NC scans and generated ones from the cohort B are used to prove the robustness; and (e) application of the GANCMLAE model in AD and MCI