The past decade has seen a substantial increase in demand for high-quality, controllable synthetic data for training medical imaging foundation models. Current synthetic Optical Coherence Tomography (OCT) data generation methods often face a trade-off between physical realism and biological congruence. To address these limitations, we propose a novel pipeline that synergizes data-driven analysis with a physics-based simulation. Our method first analyzes real OCT scans using a zero-shot Segment Anything Model (SAM) to extract realistic parameters of anatomical structures. Subsequently, these parameters are used to generate varied, biologically congruent structural masks. We then simulate a realistic distribution of optical scatterers within these masks and perform a virtual OCT scan that replicates the physics of light-tissue interaction. This physics-based generation pipeline is fundamentally different from physics-informed neural networks. We demonstrate that a MedSAM model trained exclusively on our synthetic data achieves Dice scores of 96% (Noise), 83% (Epidermis), and 97% (Dermis) on unseen real scans. This performance closely matches that of a model engaged in few-shot learning on real data (98%, 85%, and 97% respectively), validating the efficacy of our approach for training models in data-scarce scenarios.