Author: Beth Bradshaw Ghavidel, Chih-Wei Chang, Yuan Gao, Priyanka Kapoor, Shaoyan Pan, Junbo Peng, Richard L.J. Qiu, Jill Remick, Justin R. Roper, Zhen Tian, Xiaofeng Yang 👨🔬
Affiliation: Whinship Cancer Institute, Emory University, Emory University, University of Chicago, Department of Radiation Oncology and Winship Cancer Institute, Emory University 🌍
Purpose: Current cone-beam computed tomography (CBCT) typically requires no less than 200 degrees of angular projections, which prolongs scanning time and increases radiation exposure. To address these issues, this study develops a deep learning-based method for rapid 3D CBCT reconstruction utilizing orthogonal x-ray projections, specifically designed to enhance the efficiency of patient setup for gynecological cancer radiotherapy.
Methods: We introduced a cycle-domain geometry-integrated denoising diffusion probabilistic model (CG-DDPM), which achieves high-fidelity CBCT reconstruction from a single pair of orthogonal-view x-ray projections. The CG-DDPM leverages deep generative modeling, geometric transformations, and cross-domain learning to facilitate ultra-sparse CBCT reconstruction, enhancing anatomical accuracy while reducing artifacts. A key component, the Geometric Transformation Module (GTM), enhances model robustness by incorporating strict conditioning and imaging system priors. Employing a Swin-VNet-based architecture with a Cycle-domain Geometry-Integrated strategy, this approach promotes efficient information exchange between the 2D projection and 3D image domains. The CG-DDPM was evaluated using data from 22 gynecological cancer patients. Orthogonal-view x-ray projections were simulated at a source-to-detector distance of 1500 mm and resolution 768×1024 pixels. Accuracy was assessed using the following metrics: mean absolute error (MAE) in Hounsfield Units (HU), structural similarity index (SSIM), and peak signal-to-noise ratio (PSNR).
Results: Synthetic CBCT image reconstruction yielded an MAE of 30.41±10.66 HU, SSIM of 0.89±0.12, and PSNR of 28.16±8.55 dB. The proposed method demonstrated rapid 3D CBCT reconstruction with excellent image quality, crucial for CBCT-guided gynecological cancer radiotherapy.
Conclusion: The proposed CG-DDPM framework enhanced volumetric CBCT reconstructions from orthogonal x-ray projections. Integrating x-ray imaging physics and geometric priors, it markedly reduced uncertainty and artifacts. This advancement in ultra-sparse CBCT imaging expedites 3D imaging for patient setup, minimizing radiation exposure. Consequently, it reduces setup times, decreases potential patient movement, and increases comfort, thus improving setup efficiency for gynecological cancer radiotherapy.