Diffusion Model

Diffusion models perform data generation by simulating the gradual evolution of data distribution from a basic (e.g., Gaussian) to the target complex form via discretization of continuous stochastic differential equations. Diffusion models can be potentially useful in the healthcare domain, such as medical imaging synthesis/augmentation, image restoration, drug discovery, predictive modeling, etc. At CAMCA, we have conducted extensive research on the topic of diffusion models in both algorithm development and application in healthcare, including: 1) Investigating novel schemes for leveraging score functions in diffusion models (https://arxiv.org/abs/2304.08384) ; 2) More effectively and efficiently reversing the diffusion process (https://arxiv.org/abs/2302.02398) ; 3) Using diffusion probabilistic models (DDPM) for medical image analysis tasks such as denoising and image reconstruction (https://arxiv.org/abs/2209.06167).

Algorithm flow of the score function-based image denoising. The first step is to estimate the score function s(y) and the second step is to solve the equation.​

Related publications:

https://link.springer.com/chapter/10.1007/978-3-031-16446-0_62