Counterfactual Interventions

Across diabetic retinopathy, stroke, and Alzheimer’s disease, tortuosity interventions produced positive disease-probability shifts. Matched pixel-drop controls were used as a null comparison by deleting mask pixels and rerunning skeletonization and fitting. The paired scoring protocol measured disease-probability differences between perturbed and baseline generated images from the same start. Tortuosity interventions preserved segment endpoints, keeping connectivity and chord length invariant. The interventions targeted tortuosity, arc length, and radius because they were supported by observational rankings, anatomical literature, and dose-response testing needs.