Fig. 5. Validation accuracy (Acc) and AUC score for ADNI disease prediction under two scenarios.

(a). MDANN trained with cross-entropy loss. (b). MDANN trained with minimax loss. Three biased attributes: age, handedness, and educated years, were used as three embedded adversarial components (ACs) for bias mitigation. The baseline was the regular machine learning model trained without any AC. All images have been fed to a pre-trained CAE with two convolutional layers (with deep representations size of 32×32×256).