BTECF uses Bézier trees for causal retinal vessel disease analysis

Tan Su, Ethan Elio Meidinger, Lin Gu, and Ruogu Fang propose BTECF, a novel framework that overcomes a key limitation in generative counterfactuals: the inability to isolate explicit anatomical structures in retinal vessels. By encoding vessel networks as interconnected cubic-Bézier segments, BTECF preserves topological structure while allowing atomic perturbations. Coupled with a diffusion-based generator, it performs targeted interventions on geometric axes like tortuosity and caliber, keeping background fundus textures unchanged. This ensures that changes in classifier predictions are causally linked to vessel geometry, not out-of-distribution artifacts.

BTECF was validated on three distinct diseases: diabetic retinopathy, ischemic stroke, and Alzheimer's disease. A matched pixel-drop control showed that dose-responsive shifts in classifier predictions were an order of magnitude stronger than those from artifact-driven changes, proving causal isolation. This unified generative paradigm enables hypothesis verification across systemic diseases using a single anatomical biomarker. The authors demonstrate that isolated counterfactual interventions produce dose-responsive shifts, ruling out spurious correlations. Code will be released upon acceptance, paving the way for reproducible causal analysis in medical imaging.