MAGIC
Motion-aware Gaussian Splatting in Dynamic CBCT Imaging

Yufu Zhou1 Hua Chen23 Yi Liu1 Ziheng Deng1 Jun Zhao1

1 School of Biomedical Engineering, Shanghai Jiao Tong University 2 School of Medicine, Shanghai Jiao Tong University 3 Department of Radiation Oncology, Shanghai Chest Hospital

Video

Reconstruction results in the simulated regular respiration

Reconstruction results in the simulated irregular respiration

Reconstruction results in one clinical scan on Varian Truebeam system at Shanghai Chest Hospital

In the simulated studies, MAGIC consistently outperforms the comparison methods under both regular and irregular respiratory patterns, achieving superior artifact suppression and more accurate motion estimation. Notably, MAGIC excels in restoring the edges of the diaphragm and tumors, while effectively preserving fine anatomical details such as small vessels.
Although MAGIC is primarily designed for lung motion estimation and exploits the quasi-periodic nature of respiration, it demonstrates unexpected potential for cardiac motion estimation in the clinical case. This capability is evidenced in the sagittal view, by the higher-frequency cardiac motion (about 1.5 Hz) observed on the upper-left side of the diaphragm motion (about 0.3 Hz). The feasibility of dynamic cardiac reconstruction with MAGIC will be further investigated in future work.

Abstract

Dynamic cone-beam computed tomography (CBCT) provides time-resolved visualization of respiratory motion, which is crucial for precise target localization in the lung radiotherapy. While recent implicit neural representation (INR) methods have shown promise in modeling continuous spatiotemporal scenes, they are often computationally expensive. In this work, we propose a novel dynamic CBCT reconstruction framework, Motion-Aware Gaussian splatting In dynamic CBCT imaging (MAGIC), for accurate and efficient reconstruction. Instead of implicit neural networks, MAGIC represents the dynamic scene using learnable Gaussian primitives, which reduces reconstruction time and maintains representation flexibility. We introduce a motion-aware decomposition (MADE) method for Gaussians to decouple static anatomical components and motion variations. The decomposition allows the network to focus its capacity on the deformation of dynamic regions, which facilitates more accurate motion modeling and shortens reconstruction time. Furthermore, we design a respiratory deformation network (RED) with an encoder of ten feature planes (Deca-Planes) to effectively capture inter- and intra-cycle similarities from quasi-periodic respiration, thus improving motion estimation accuracy. Experiments on simulated data, a physical phantom and clinical data demonstrate that compared to other methods, MAGIC achieves higher image quality and motion fidelity while accelerating reconstruction. These results indicate that MAGIC offers an effective and efficient alternative for dynamic CBCT imaging in the respiration-related radiotherapy.