New framework generates synthetic brain scans with known causal ground truth

A team led by Eryn Libert-Scott and colleagues at the University of Calgary has introduced a neuroimaging simulation framework that creates synthetic brain MRIs with fully controllable causal structures. The method works by sampling latent anatomical variability from a subspace learned from real T1-weighted images, then deforming a template to generate unique simulated subjects. Causal relationships are encoded by precisely altering volumes of any region of interest without introducing unwanted global artifacts, achieving relative volume errors of just 0.3-2.66% for targeted regions and mean absolute errors of 0.034-0.397 ml for non-target areas. This level of control allows researchers to know exactly which causal relationships exist between variables like disease factors and brain morphology. The paper, submitted to the Journal of Biomedical and Health Informatics, also includes an initial evaluation of existing causal discovery methods on these synthetic datasets, revealing that current approaches have limited ability to suppress spurious connections, underscoring the need for image-appropriate causal AI techniques. The framework code is publicly available, providing the missing ground-truth data needed for objective benchmarking.