Image: Researchers are improving generative AI models for real-world medical imaging (Photo courtesy of 123RF)
Diffusion models are a type of deep generative models that are extremely successful in applications like image generation and audio synthesis, as well as medical imaging and molecule design. Diffusion models are designed to learn the data distribution, which is important for deciphering large-scale and complex real-world data. Presently, there are several limitations regarding the practical applications of diffusion models. For instance, the training and inference of diffusion models are both data-intensive and computationally demanding, which limits their usage across various scientific disciplines. The images generated in real-world medical imaging are always high-resolution and high-dimensional, far beyond what can be managed by existing diffusion models in terms of memory and time efficiency. Additionally, diffusion models have an undesirably lengthy inference time due to the iterative sampling procedure.
The Michigan Engineering research team at the University of Michigan (Ann Arbor, MI, USA) is working on developing new and more efficient diffusion models that can surpass the current limitations. The team is focusing on examining how diffusion models can be applied to inverse problems, which is when a set of observations are utilized to determine the factors that generated the results. The team is working to improve the practical applicability and mathematical interpretability of diffusion models by developing new architecture designs and latent embeddings.
The researchers are also developing new techniques to improve the training and sampling efficiency of diffusion models. They are working to create computationally efficient diffusion models for high-dimensional data that could further improve data, memory, and time efficiency. This could significantly improve applications such as high-dimensional, high-resolution biomedical imaging, as well as motion prediction based on high-dimensional dynamic imaging.
“Generative models are one of the hottest topics in machine learning right now, and I’m excited to have the opportunity to investigate their potential for solving inverse problems, especially in medical imaging,” said Fessler, the William L. Root Collegiate Professor of EECS. “We’re hoping to apply the methods developed in this project to large-scale 3D medical imaging applications, like low-dose X-ray CT and accelerated MRI.”
University of Michigan