Artemis: New Causal AI Framework Cleans Brain Imaging Confounders
Demographic biases in brain network analysis? Artemis intervenes at the region level.
A team of researchers from multiple universities (including University of Pittsburgh, USC, and others) has released Artemis (Anatomy-Resolved inTervention for Eliminating Multimodal NeuroImage confounderS) on arXiv. The framework addresses a critical flaw in brain network analysis: demographic factors such as age and sex systematically confound the relationship between connectivity and clinical outcomes. Current GNN-based methods often exploit these spurious correlations rather than learning causally invariant representations, especially since brain regions show varying sensitivity to demographic factors.
Artemis introduces a region-level causal intervention mechanism that learns confounder representations independently for each brain region using lightweight parameters. It integrates both functional connectivity (from fMRI) and structural connectivity (from DTI) as a plug-in module compatible with any GNN backbone. On three benchmarks—ADNI for Alzheimer's disease diagnosis, OASIS for dementia staging, and HCP for sex classification—Artemis achieved consistent improvements over representative GNN baselines. Additional experiments confirmed statistical significance and neuroscientific interpretability, making it a practical tool for clinical neuroimaging analysis.
- Artemis performs per-brain-region causal intervention to remove demographic confounders (age, sex) from multimodal brain network data.
- Achieved consistent accuracy gains on three benchmarks: ADNI, OASIS, and HCP, outperforming standard GNN baselines.
- Integrates both fMRI functional connectivity and DTI structural connectivity as a drop-in module for any GNN backbone.
Why It Matters
Enables more reliable AI-driven diagnosis of neurological conditions by removing biases that skew brain network analysis.