Author: Hania A. Al-Hallaq, Chih-Wei Chang, Anees H. Dhabaan, Yuan Gao, Shaoyan Pan, Junbo Peng, Richard L.J. Qiu, Keyur Shah, Sibo Tian, Zhen Tian, Xiaofeng Yang, David Yu, Jun Zhou 👨🔬
Affiliation: Emory University, Whinship Cancer Institute, Emory University, University of Chicago, Department of Radiation Oncology and Winship Cancer Institute, Emory University 🌍
Purpose: Traditional cone-beam computed tomography (CBCT) often requires multiple angular projections, increasing radiation exposure and extending scanning times, which may lead to heightened patient discomfort and complicate treatment setups. This study presents a deep learning-based method for rapid 3D volumetric reconstruction using two orthogonal X-ray projections, aiming to enhance patient setup efficiency for head and neck (H&N) cancer radiotherapy while reducing radiation dose and scan duration.
Methods: We developed a cycle-domain denoising diffusion probabilistic model (CDDPM), that achieves high-fidelity CBCT reconstruction from just two orthogonal-view X-ray projections. The model incorporates a geometric transformation module that enhances robustness by enforcing strict conditioning and integrating imaging system priors. Employing a Swin-Vnet-based architecture with a cycle-domain geometry-integrated strategy, this method facilitates effective information transfer between projection and image domains. The evaluation utilized 28 H&N cancer patients, employing simulated orthogonal-view X-ray projections with a resolution of 768×1024 pixels and a source-to-detector distance of 1500 mm. Key metrics assessed included mean absolute error (MAE) in Hounsfield units (HU), structural similarity index (SSIM), and peak signal-to-noise ratio (PSNR).
Results: The proposed model achieved a MAE of 31.48 ± 12.36 HU, a SSIM of 0.79 ± 0.15, and a PSNR of 27.26 ± 9.45 dB, demonstrating the capability of providing high-quality images, an alternative to CBCT-guided H&N cancer radiotherapy.
Conclusion: The proposed CDDPM framework substantially enhances volumetric CBCT imaging by utilizing ultra-sparse X-ray projections, integrating X-ray imaging physics, and geometric priors to effectively reduce uncertainties and artifacts. By leveraging deep generative modeling, geometric transformations, and cross-domain learning, CDDPM enables ultra-sparse 3D reconstruction, improving anatomical accuracy and reducing artifacts. This innovation in ultra-sparse CBCT imaging accelerates 3D imaging for patient setup, substantially reduces radiation exposure, and improves patient comfort, thereby increasing the efficiency of H&N cancer radiotherapy.