Impact of Adversarial-Cooperative Architecture on Alzheimer's Diagnosis and Confound Regression

This study investigates the performance of a dual adversarial-cooperative deep neural network in diagnosing Alzheimer's disease (AD) from 3D MRI neuroimaging data under various training configurations. Neuroimaging datasets often contain confounding variables, potentially biasing model predictions and affecting fairness. We systematically examine how different strategies for confound label masking and confound loss weighting influence diagnostic accuracy and the model's ability to mitigate confounding effects. Across 190 experiments, the model achieves up to 76.6% accuracy in AD detection while maintaining a low overall Generalized Entropy Index (GEI) of 0.11, suggesting moderate individual-level fairness. However, some configurations that more effectively suppressed race/ethnicity predictions exhibited slightly less favorable GEI, revealing a trade-off between confound suppression and fairness. These findings suggest that imaging features linked to group identity may overlap with disease-relevant patterns, making it difficult to fully eliminate their influence. Our results highlight the need for careful, context-aware approaches to fairness and confound mitigation in medical imaging.