Home | Publications | FDT+26

Structured Loss Amplification for U-Net-Based Human-Scale Dark-Field CT Streak Reduction

MCML Authors

Abstract

Streak artifacts remain a substantial challenge for the clinical translation of human-scale dark-field computed tomography. Supervised learning approaches using convolutional neural networks have shown promise in artifact suppression, but typically omit domain-specific prior knowledge. In this work, we present a structured loss formulation that integrates spatial and frequency-informed bias into U-Net-based streak reduction by selectively amplifying pixel-wise loss, guiding the model towards high-fidelity predictions during training. Performance was analyzed across independently trained models and quantified via contrast-to-noise ratio (CNR) and full-widthat-half-maximum (FWHM) of the line spread function at the air-tissue interface. As a substitute for clinical patient data, test metrics were computed from a scan of ventilated ex vivo porcine lungs. Using center- and high-frequency amplification, CNR was enhanced by 24% for L1 and 33 % for L2 loss while preserving adequate image resolution. These results motivate the use of structured loss amplification as a flexible design choice for neural network architectures and demonstrate its effectiveness for advanced streak reduction in dark-field imaging.

inproceedings FDT+26

CT Meeting 2026

9th International Conference on Image Formation in X-Ray Computed Tomography. Salt Lake City, USA, Jun 01-03, 2026. To be published. Preprint available.

Authors

D. Frey • T. Dorosti • J. B. Thalhammer • J. F. Hilmer • P. Bleuel • T. Hiu • S. Peterhansl • J. McGinnis • T. Koehler • D. Pfeiffer • F. Pfeiffer • D. Rückert • F. Schaff

Links

PDF

Research Area

 C1 | Medicine

BibTeXKey: FDT+26

#p-rueckert #p-wiestler
Back to Top