Inpainting the Sinogram from Computed Tomography using Latent Diffusion Model and Physics
Keywords: Sinogram Inpainting, Physics, Latent Diffusion Model, X-ray Imaging
Abstract: Computed Tomography (CT) is a widely used non-invasive imaging technique for materials at microscopic or sub-microscopic length scales in synchrotron radiation facilities. Typically, the object is rotated relative to the X-ray beam, and 2D projection images are recorded by the detector at different rotation angles. The 3D object is then reconstructed by combining these projections and solving a computationally demanding inverse problem. The quality of the reconstructed image is critical for scientific analysis and is influenced by various factors, including the number of projections, exposure time or dose, and the reconstruction algorithm. In this work, we develop a foundation model by integrating a Generative AI-based Latent Diffusion Model (LDM) with physics-based domain knowledge. Specifically, we first incorporate a set of loss functions into our LDM that accurately capture the physical properties of the CT data acquisition process. We demonstrate that addition of these loss functions aids in stable training of the autoencoder in the LDM and improves its accuracy. The autoencoder and the Diffusion model of the LDM is trained with real-world experimental data. Collecting real world experimental data from Synchrotron beamlines is often time-consuming and challenging. We demonstrate that the autoencoder trained with a combination of real world experimental data and phantom shapes features also performs comparable to the autoencoder trained with real world data. Second, we introduce a novel image blending method to combine the LDM’s generated output with the original, extremely sparse sinogram data. Since our model integrates physics-guided loss functions focused on CT data acquisition, it simplifies the creation of downstream tasks and facilitates the adaptation of new features from different experiments. Our experimental evaluation demonstrates improvements of upto 23.5 % in SSIM for sinogram quality and 13.8 % for reconstructed image quality compared to state-of-the-art techniques.
Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 13542
Loading