Stochastic Flow Inference for Medical Image Digital Twins

Digital twins in medicine offer personalized models that replicate patient-specific anatomy and physiology, enabling accurate diagnosis, individualized treatment planning, and predictive simulations. However, constructing reliable digital twins, particularly for complex anatomical structures like the human cochlea, requires robust generative modeling methods capable of synthesizing high-resolution medical images. In this paper, we propose a novel Langevin Variational Autoencoder (LVAE) based generative framework, integrating stochastic flows to enhance the expressiveness and computational efficiency of variational inference. By employing a quasi-symplectic Langevin flow, the proposed LVAE overcomes the limitations of existing approaches, reducing computational overhead associated with Hessian computations. Experimental evaluations on a large-scale cochlea CT dataset demonstrate superior performance of the LVAE in terms of inference accuracy, convergence speed, and image realism compared to existing stochastic flow-based VAEs. Our approach provides an effective foundation for building clinically reliable cochlear digital twins, thereby advancing personalized medicine and improving clinical outcomes for patients with hearing impairments.