Speaker: Dufan Wu, PhD | Assistant Professor, Department of Radiology, Massachusetts General Hospital and Harvard Medical School
Date: Tuesday, November 12th, 2024
Time: 11:00 AM Central Time
Location: Zoom
Title: Diffusion Bridge Models for CT Image Restoration and Reconstruction
Abstract: Generative diffusion models have shown great ability to generate high-quality images and have been investigated for image reconstruction, but the slow inference speed is one of the major limitations of the diffusion models. Despite various acceleration techniques, such as higher-order solvers and distillation, starting the inference from white Gaussian noise limits the efficiency of conventional diffusion models in image restoration problems. Diffusion bridge models were proposed to tackle this challenge, where the stochastic differential equations (SDE) were built to connect the degraded image and the original image directly, greatly enhancing the inference efficiency via a much closer starting point, achieving up to 10 times acceleration compared to conventional diffusion models.
This talk will introduce our recent work on developing diffusion bridge models for CT image restoration and reconstruction. Inspired by DDIM, we proposed a non-Markovian chain-based diffusion bridge model for image restoration, “Implicit Image-to-Image Schrodinger Bridge (I3SB)”, achieving a 2 to 4 times acceleration compared to vanilla bridge models. To improve data fidelity and reduce hallucination, we proposed a novel method to incorporate data constraint into the reverse sampling of diffusion bridge models in a cost-function framework. Applications including denoising, sparse-view reconstruction, and motion correction were used to demonstrate the efficacy of the methods.
Bio: Dr. Dufan Wu is an Assistant Professor of Radiology in the Center for Advanced Medical Computing and Analysis, Department of Radiology, at Massachusetts General Hospital (MGH) and Harvard Medical School (HMS). He received his PhD in Nuclear Science and Technology from the Department of Engineering Physics at Tsinghua University in 2016 and has been working at MGH since then. His main research interests include AI in medical image processing and reconstruction, photon counting and spectral CT, and translational research of AI in medical image analysis.