09/21/2026
By Marley O'Neil

The Francis College of Engineering, Department of Electrical and Computer Engineering, invites you to attend a doctoral dissertation defense by Changsheng Fang titled “Probabilistic Generative Modeling for Incomplete CT Reconstruction.”

  • Date: Monday, Sept. 28, 2026
  • Time: 10 a.m. – noon
  • Location: Ball Hall 302

Committee

  • Advisor: Hengyong Yu, professor, Electrical and Computer Engineering, University of Massachusetts Lowell
  • Yan Luo, professor, Electrical and Computer Engineering, University of Massachusetts Lowell
  • SeungWoo Son, associate professor, Electrical and Computer Engineering, University of Massachusetts Lowell
  • Dayang Wang, Ph.D., research scientist, Subtle Medical Inc.

Brief Abstract

Computed tomography (CT) is widely used in clinical diagnosis and intervention, but high-quality reconstruction typically requires sufficient projection measurements. In practice, complete data acquisition is often limited by radiation dose, scanning time, system geometry, patient motion, clinical workflow, and other constraints. Therefore, incomplete CT reconstruction, including sparse-view CT and limited-angle CT, remains a challenging ill-posed inverse problem that can cause severe artifacts, structural distortion, and loss of anatomical details.

This dissertation investigates probabilistic generative modeling methods for accurate, efficient, and physically consistent incomplete CT reconstruction. First, a Residual Poisson Flow Generative Model (ResPF) is proposed for sparse-view CT reconstruction. By adapting Poisson Flow Generative Models to paired sparse-view and full-view CT data, ResPF learns deterministic generative trajectories for high-quality image recovery. Accelerated sampling, projection-domain data consistency, and residual fusion are incorporated to improve reconstruction fidelity and reduce inference time.

Second, a Lightweight Wavelet Diffusion with Kolmogorov–Arnold Network (WDK-Net) is developed for limited-angle cardiac CT reconstruction. WDK-Net uses wavelet-domain diffusion to recover global anatomical structures in a reduced-dimensional space, high-frequency enhancement to restore directional details, and Kolmogorov–Arnold Network-based refinement to suppress artifacts and improve local structures. This structure–detail decoupled design improves reconstruction quality in multiple limited-angle settings while maintaining practical efficiency.

Third, a data-consistency-inspired conditional diffusion framework (DCI-diff) is proposed for limited-angle CT reconstruction. Unlike conventional diffusion methods that employ uniform training objectives and enforce data consistency only during sampling, the proposed framework explicitly incorporates the heterogeneous physical constraints induced by incomplete angular coverage into both training and inference. By integrating uncertainty-aware spatial modulation, global structural consistency, and physically constrained sampling, the method improves reconstruction stability, suppresses structural error propagation in ill-posed regions, and achieves more faithful recovery under severe missing-angle conditions.

The proposed methods are evaluated using both simulation and clinical CT datasets to assess reconstruction accuracy, artifact suppression, structural preservation, and computational efficiency under different incomplete acquisition settings. Experimental results demonstrate that probabilistic generative modeling, when jointly designed with imaging physics, uncertainty modeling, and domain-specific reconstruction strategies, provides an effective and practical solution for incomplete CT reconstruction.