Adaptive Diffusion Priors on Pre-clinical CT Reconstruction

Theresa Hiu, Daniel Frey, Tina Dorosti, Johannes Thalhammer, Sebastian Peterhansl, Zijin Huang, Simon Zandarco, Franz Pfeiffer, Florian Schaff

Keywords: Computed tomography, dark-field imaging, diffusion models

TL;DR: The method combines a pretrained diffusion prior with physics-based consistency and incorporates low-rank adaptation (LoRA) to adapt the pretrained diffusion model to the measured data for pre-clinical dark-field micro-CT reconstruction.

Registration Requirement: Yes

Abstract: X-ray dark-field imaging enables visualization of lung microstructure and the detection of pulmonary diseases. However, in pre-clinical micro-CT studies, the reconstruction suffers from severe streak artifacts due to highly undersampled acquisitions constrained by radiation dose. We leverage a diffusion-based reconstruction framework by extending the Deep Diffusion Image Prior (DDIP) to dark-field CT. The method combines a pretrained diffusion prior with physics-based consistency and incorporates low-rank adaptation (LoRA) for improved robustness to out-of-distribution data. Experiments demonstrate increased contrast-to-noise ratio and artifact suppression while preserving edge sharpness, enabling higher-quality dark-field imaging for dose-constrained longitudinal studies.

Visa & Travel: No

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 114