Translation as a Reference for One-Step OCT-to-OCTA Generation

Optical coherence tomography (OCT) provides volumetric retinal structure, whereas OCT angiography (OCTA) visualizes retinal microvasculature through flow-sensitive contrast. Synthesizing OCTA from routinely acquired OCT can support vascular assessment when OCTA measurements are unavailable or unreliable. Despite recent progress, deterministic translation may underrepresent angiographic signals not fully specified by OCT, whereas direct full-target generation allows the complete OCTA volume to vary, including spatial organization that should remain tightly paired with the input. We hypothesize that using translation as a reference and reserving generation for residual angiographic variation simplifies one-step OCT-to-OCTA synthesis. In this paper, we introduce reference-centered generation and instantiate it with Pixel MeanFlow by decomposing each prediction into a subject-specific OCTA reference and a generated residual. The method requires neither vessel nor retinal-layer annotations and achieves state-of-the-art performance on OCTA-3M and OCTA-6M. Mechanism analysis further supports the hypothesis that reference centering simplifies the learned finite-interval transport.