SONAR: A Physics-constrained Neural Representation for X-ray Dark-field CT

Daniel Frey, Theresa Hiu, Julian McGinnis, Tina Dorosti, Johannes Thalhammer, Sebastian Peterhansl, Zijin Huang, Franz Pfeiffer, Daniel Rueckert, Florian Schaff

Keywords: Computed tomography, dark-field imaging, implicit neural representations

TL;DR: SONAR is a neural representation adaptively trained under Talbot–Lau interferometer physics to suppress streak artifacts in X-ray dark-field CT.

Registration Requirement: Yes

Abstract: Dark-field computed tomography (DFCT) enables functional lung imaging with small-angle X-ray scattering, but reconstructions are often degraded by streak artifacts. We propose SONAR (Shot-Optimized Neural Adaptive Representation), a projection-based implicit neural representation (INR) that jointly models transmission, phase shift, and dark-field signals across neighboring shots using a physics-based Talbot–Lau interferometer forward model. By leveraging adaptive per-projection optimization, SONAR effectively enables stabilized phase retrieval and suppresses streak artifacts for improved DFCT image quality on a grating-based human-scale prototype.

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 112