Home  | Publications | FDM+26

Robust Sparse-View Dark-Field CT With 3D Gaussian Splatting

MCML Authors

Abstract

INTRODUCTION: Streak artifacts constitute a major challenge in X-ray darkfield computed tomography (DFCT). Convolutional neural networks have shown promising results for streak reduction, but they rely on dedicated ground truth data with limited availability [1]. In contrast to post-processing approaches, this work aims to prevent streak formation at the reconstruction stage using 3D Gaussian splatting (3DGS).<br>METHODS: 3DGS was implemented using R2-Gaussian, as described in [2], and sparsely acquired dark-field projections reconstructed in comparison to Feldkamp–Davis–Kress (FDK). Reconstruction quality was assessed based on peak signal-tonoise ratio (PSNR) and structural similarity index measure (SSIM) in reference to full-view FDK (2324 projections).<br>RESULTS AND CONCLUSION: Visual and quantitative evaluation highlight the superior performance of 3DGS compared to FDK, as illustrated in Fig. 1.<br>Due to imperfections in the full-view FDK reference, metrics can deviate from the visual impression. SSIM provides a more reliable assessment owing to its larger noise resilience compared to PSNR. Our findings underscore the potential of 3DGS to enable streak-free DFCT reconstruction.

inproceedings FDM+26

ISBI 2026

IEEE 23rd International Symposium on Biomedical Imaging. London, UK, Apr 08-11, 2026. To be published. Preprint available.

Authors

D. Frey • T. Dorosti • J. McGinnis • T. Hiu • F. I. Ozlugedik • J. B. Thalhammer • S. Peterhansl • D. Rückert • F. Pfeiffer • F. Schaff

Links

PDF

Research Area

 C1 | Medicine

BibTeXKey: FDM+26

#p-rueckert #p-wiestler
Back to Top