Geometry-Embedded Neural Networks Cone Beam CT Reconstruction for Arbitrary Scanning Trajectories

Nele Blum, Max Stickel, Moritz Schaar, Thorsten Buzug, Maik Stille

Keywords: CT Reconstruction, Deep Learning, Non-Circular Scanning Trajectories

Registration Requirement: Yes

Abstract:

Flexible scanning trajectories for CBCT systems offer enhanced imaging capabilities but
pose challenges for conventional reconstruction algorithms. We present a novel deep learning-
based reconstruction framework that explicitly integrates trajectory geometry information
through differentiable back projection operators embedded within the network architecture.
Our approach comprises three cascaded networks: a preprocessing network for projection
filtering, an encoder-decoder reconstruction network with multi-scale back projection opera-
tors, enabling transition from projection to image domain, and a postprocessing network for
final refinement. Evaluation on 2,000 test objects demonstrates substantial improvements
over conventional filtered back projection (FBP), achieving an MSE of 4.8e−3 (vs.2.4e−2
for FBP) and SSIM of 0.96 (vs. 0.54 for FBP).

Visa & Travel: No

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 63