Deep-Learning-based Arbitrary Trajectory CBCT Reconstruction

• Applicant: Professor Dr.-Ing. Andreas Maier
• Subject Area: Medical Physics, Biomedical Technology
• Term: since 2025
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 569542899

This research project aims to significantly improve the quality and speed of image reconstruction for arbitrary CT trajectories by developing a novel alternative to computationally intensive iterative methods. The approach centers on three core innovations: first, a compressed redundancy weight modeling technique to achieve low-dimensional representations that enhance convergence and generalization across diverse CT configurations; second, a fast shift-variant filtered-backprojection framework building upon and extending Defrise and Clack’s methodology to improve efficiency in both training and reconstruction phases; and third, a neural network-based reconstruction model capable of learning shift-variant redundancy patterns, enabling fast, artifact-free reconstruction tailored to specific CT trajectories. Collectively, these contributions will provide a robust and versatile foundation for next-generation CT imaging. The project also sets the stage for a second phase, which will explore end-to-end differentiable models for physical artifact correction—such as scatter and beam hardening—and direct optimization of CT trajectories, further enhancing imaging performance. All developed code will be released as open source.

• DFG Programme: Research Grants
• International Connection: USA
• Co-Investigators: Dr.-Ing. Siming Bayer; Professor Dr.-Ing. Gabriel Herl
• Cooperation Partner: Professor Adam Wang, Ph.D.