Artemis Enhances Neuroimaging AI by Removing Demographic Confounders

Analyzing brain networks using multimodal neuroimaging data, which combines fMRI and DTI, is a powerful tool for understanding brain function and structure. However, demographic variables such as age and sex often act as confounders, causing Graph Neural Networks (GNNs) to learn spurious correlations rather than true causal relationships between brain connectivity and clinical outcomes. Existing causal GNN methods often lack domain-specific mechanisms to address these real-world confounders in neuroimaging. Furthermore, different brain regions exhibit varying sensitivities to demographic factors, necessitating a region-aware adjustment approach. Artemis introduces a novel region-level causal framework that intervenes at each brain region independently. It learns specific confounder representations for each region using lightweight parameters, effectively adjusting for these biases. This plug-in module is compatible with various GNN backbones and significantly improves performance on disease diagnosis, dementia staging, and sex classification benchmarks, while also offering neuroscientific interpretability.