Bézier Tree Encoding Counterfactual Framework

Some parts of BTECF are disease-agnostic, while LoRA, ControlNet, and the classifier are retrained per disease cohort. BTECF contains six modules spanning segmentation, Bézier parameterization, generation, classification, and paired counterfactual scoring. BTECF is designed to represent retinal vascular trees with cubic Bézier segments whose topology is preserved and whose geometric axes can be independently perturbed. BTECF uses the same Bézier encoding for observational feature analysis and interventional counterfactual generation.