Computed Tomography From a Few Projections
Event details
| Date | 14.06.2024 |
| Hour | 09:00 › 11:00 |
| Speaker | Zhantao Deng |
| Location | |
| Category | Conferences - Seminars |
EDIC candidacy exam
Exam president: Prof. Maria Brbic
Thesis advisor: Prof. Pascal Fua
Co-examiner: Prof. Graham Knott
Abstract
Computed Tomography (CT) is widely used in material and medical applications. However, conventional CT reconstruction algorithms have dedicated drawbacks. For example, the filtered back projection (FBP) method requires CT projections to be complete and continuous to have ideal reconstruction; the iterative reconstruction (IR) method is computationally heavy and slow. The fast-developing deep learning (DL) researches show the potential of acquiring high-quality CT reconstruction from challenging scenarios such as sparse view (SV) CT and limited angle (LA) CT. But the instability and lack of generalization of conventional DL methods restrict their application in complex real scenarios. Recently emerged diffusion models (DM) demonstrate a strong capability of preserving high-quality prior knowledge from data, which can serve as the regularization term in conventional CT reconstruction problems and drastically improve performance. However, there are still several limitations in current DM pipelines for CT. Firstly, the DMs are trained on general CT datasets with post-processed or synthesized projection data which impacts the generality of the models, because models lose the opportunity to learn the potentially helpful priors in the filtered-out data and real data. Besides, current state-of-the-art methods combine the conventional IR with DM by inserting the IR process into the diffusion process, leading to long iterations and slow. Then, although the DM pipelines outperform conventional IR and DL methods in the SV and LA reconstruction task, in more extreme scenarios, their performance is degraded because of limited available projections and the violation of the basic assumption regarding the noise distribution in IR pipelines. Throughout this project, we will explore the solutions to the limitations above by: Building a better IR-based DM pipeline adaptive to noise under different conditions; Accelerating the IR-based DM pipeline while keeping its flexibility and stability; Integrating DM into lower data level to better leverage the data prior.
Background papers
Physics-/Model-Based and Data-Driven Methods for Low-Dose Computed Tomography (2022)
Solving 3D Inverse Problems using Pre-trained 2D Diffusion Models
Generative Diffusion Prior for Unified Image Restoration and Enhancement
Exam president: Prof. Maria Brbic
Thesis advisor: Prof. Pascal Fua
Co-examiner: Prof. Graham Knott
Abstract
Computed Tomography (CT) is widely used in material and medical applications. However, conventional CT reconstruction algorithms have dedicated drawbacks. For example, the filtered back projection (FBP) method requires CT projections to be complete and continuous to have ideal reconstruction; the iterative reconstruction (IR) method is computationally heavy and slow. The fast-developing deep learning (DL) researches show the potential of acquiring high-quality CT reconstruction from challenging scenarios such as sparse view (SV) CT and limited angle (LA) CT. But the instability and lack of generalization of conventional DL methods restrict their application in complex real scenarios. Recently emerged diffusion models (DM) demonstrate a strong capability of preserving high-quality prior knowledge from data, which can serve as the regularization term in conventional CT reconstruction problems and drastically improve performance. However, there are still several limitations in current DM pipelines for CT. Firstly, the DMs are trained on general CT datasets with post-processed or synthesized projection data which impacts the generality of the models, because models lose the opportunity to learn the potentially helpful priors in the filtered-out data and real data. Besides, current state-of-the-art methods combine the conventional IR with DM by inserting the IR process into the diffusion process, leading to long iterations and slow. Then, although the DM pipelines outperform conventional IR and DL methods in the SV and LA reconstruction task, in more extreme scenarios, their performance is degraded because of limited available projections and the violation of the basic assumption regarding the noise distribution in IR pipelines. Throughout this project, we will explore the solutions to the limitations above by: Building a better IR-based DM pipeline adaptive to noise under different conditions; Accelerating the IR-based DM pipeline while keeping its flexibility and stability; Integrating DM into lower data level to better leverage the data prior.
Background papers
Physics-/Model-Based and Data-Driven Methods for Low-Dose Computed Tomography (2022)
Solving 3D Inverse Problems using Pre-trained 2D Diffusion Models
Generative Diffusion Prior for Unified Image Restoration and Enhancement
Practical information
- General public
- Free