To address the performance bottleneck in deep learning caused by the scarcity of paired orthopedic ultrasound data, this paper proposes a hybrid data generation framework fusing physical acoustic simulation with Generative Adversarial Networks (GANs). We first utilize Monte Carlo simulation with generalized parameter sampling to construct structured synthetic data covering diverse anatomical variations. This is followed by a proposed VR-CycleGAN for Sim-to-Real domain adaptation. Unlike standard CycleGAN variants that are prone to geometric distortions, our model integrates a VAE module to explicitly decouple stochastic texture synthesis from anatomical structure. This design imparts realistic clinical textures while rigorously preserving anatomical integrity using unpaired data. Experimental results demonstrate that the generated images not only achieve superior perceptual quality but also exhibit Nakagami distribution parameters (m = 0.5856, Ω = 0.0537) that closely align with real ultrasound backscattering statistics (m = 0.5991, Ω = 0.0608). This quantitative evidence confirms the framework’s capability to reproduce complex physical speckle features and validates the potential of integrating physical priors with data-driven approaches for high-fidelity medical image synthesis.
Ruize Li (Wed,) studied this question.