Add to Calendar 2026-11-16T11:00:00 2026-11-16T13:00:00 Image-Quality Limits and Reconstruction Methods in Preclinical Cone-Beam Micro-CT Event Information: Abstract: X-ray micro-computed tomography (μCT) reconstructs a three-dimensional volume from hundreds of X-ray projections. Any reduction in dose or scan time degrades the projection data and with it the diagnostic value of the reconstruction. This thesis asks how close reconstruction methods come to the image quality that a given set of projections can support. Chapter 1 maps the field through a citation traversal and semantic analysis of about seven thousand papers. Chapter 2 establishes the ceiling of achievable image quality. It derives the photon-noise floor and a closed-form bound on resolution and detectability from noise and angular sampling, and distinguishes the system metrics (MTF, NPS, NEQ, d′) from the reference-based metrics (SSIM, PSNR) that dominate the machine-learning literature. The remaining chapters introduce and measure the different reconstruction routes from projections to volume against that ceiling on one scanner, one protocol, one image-quality phantom and one prospectively gated in-vivo mouse scan. Tuned analytic FDK serves as the reference for all later methods. Adomain-matched network that completes missing projections outperforms transferred inpainting models on every system metric. Iterative reconstruction (SIRT, PWLS, MLEM) reaches a plateau of task-equivalent solutions under this protocol, and only the total-variation prior improves detectability. Learned reconstruction through a voxel grid, a neural field and a Gaussian cloud differs in how each fills the null space that the data leave open. The neural field scores best on every reference-free metric on the phantom but produces unsupported anatomy in the animal, whereas the Gaussian cloud leaves the null space empty and achieves high soft-tissue detectability. A dynamic extension of the cloud, constrained by mass preservation and an air-sourced divergence, reconstructs both gated phases together and provides a ventilation map that agrees with an independent density measurement.A recurring result is that reference-based metrics can be misleading. Agreement in the projection domain does not guarantee a faithful reconstruction, and routine post-processing such as tile blending can change diagnostic metrics without adding information. The contributions, published across several papers, preprints and public tools, support a physics-based standard of evaluation for learning-based μCT reconstruction that states the achievable ceiling and reports the system metrics alongside every reference-based score. Event Location: HENN 318