10/26/2026
By Marley O'Neil

The Francis College of Engineering, Department of Electrical and Computer Engineering, invites you to attend Haodong Li’s doctoral dissertation proposal defense titled "Toward Generalizable Medical Image Reconstruction with Physics-Guided Diffusion Priors."

Details

  • Defense date: Friday, Nov. 6, 2026
  • Time: 10 – 11:30 a.m.
  • Location: Ball Hall 302

Committee Members

  • Advisor: Hengyong Yu, Ph.D., Electrical and Computer Engineering Department, UMass Lowell
  • Tricia Chigan, Professor, Electrical and Computer Engineering, UMass Lowell
  • Yan Luo, Professor, Electrical and Computer Engineering, UMass Lowell
  • Seung Woo Son, Associate Professor, Electrical and Computer Engineering, UMass Lowell

Abstract

Medical images are often reconstructed from incomplete or noisy physical measurements. Sparse-view computed tomography (CT) and undersampled magnetic resonance imaging (MRI) require image priors and measurement constraints. Diffusion models offer expressive priors, but their applicability can change with anatomy, scanners and acquisition conditions. This dissertation investigates physics-guided diffusion reconstruction through two complementary questions: prior applicability across image distributions and framework applicability across forward operators. Cross-Distribution Diffusion Priors-Driven Iterative Reconstruction (CDPIR) combines an image-domain diffusion transformer with residual-guided iterative CT reconstruction. Evaluations examine transfer between CT datasets and application to clinical cardiac and photon-counting CT measurements. They support applicability in the tested settings while exposing substantial computation cost and questions about prior–physics coupling. RAN-Diff addresses this organization through clean-image proposal, measurement-based anchoring and re-noising. Its operator-space analysis clarifies the assumptions behind idealized anchoring and distinguishes that analysis from practical solvers. CT and MRI implementations use separate pretrained priors within the same transition structure. Image-quality comparisons and preliminary component and timing analysis support these implementations, with conclusions bounded by their experimental scope. The proposed final stage will explore a text-conditioned Stable Diffusion prior within the decoupled CT reconstruction framework. Model selection, medical adaptation, the connection between latent inference and physical updates, and the conditions for useful guidance remain open. The six-month completion plan combines this focused exploration with further validation and integration of the existing studies.