Artemis: Anatomy-Resolved inTervention for Eliminating Multimodal NeuroImage confounderS

Short summary

Artemis tackles demographic confounding in neuroimaging by applying region-level causal intervention to Graph Neural Networks analyzing multimodal brain connectivity (fMRI + DTI). The framework learns region-specific confounder representations to prevent spurious shortcuts while remaining compatible with any GNN backbone. Validated across three clinical benchmarks (ADNI, OASIS, HCP), it demonstrates significant improvements in disease diagnosis, dementia staging, and sex classification with strong neuroscientific interpretability.

• •Region-level causal intervention addresses age and sex confounding in neuroimaging GNNs
• •Multimodal integration of functional (fMRI) and structural (DTI) connectivity as a plug-in module
• •Validated on three clinical benchmarks with consistent improvements and statistical significance

Generated with AI, which can make mistakes.