Physics-constrained dual-domain network for CBCT reconstruction from orthogonal X-rays in gynecologic radiotherapy.

Authors

Abstract

Daily cone-beam computed tomography (CBCT) is widely used for image-guided radiotherapy (IGRT) in gynecologic (GYN) cancer to verify patient setup and visualize inter-fraction pelvic anatomical variations, where volumetric soft-tissue information is critical due to substantial organ motion and deformation. However, routine CBCT acquisition increases treatment time and cumulative imaging dose over multi-week treatment courses. In clinical practice, orthogonal two-dimensional (2D) kilovoltage (kV) X-ray imaging is often used for rapid setup verification but provides limited soft-tissue information and cannot adequately capture internal anatomical changes relevant to pelvic radiotherapy. Recent deep learning approaches have shown promises for reconstructing three-dimensional (3D) images from sparse projections, yet many methods lack explicit physical constraints or fail to fully exploit the geometric coupling between projection data and volumetric anatomy, particularly in patient-specific GYN radiotherapy settings. The purpose of this study is to develop and evaluate a patient-specific framework that reconstructs volumetric CBCT images from orthogonal anterior-posterior (AP) and lateral (LAT) X-ray projections, enabling recovery of clinically relevant three-dimensional (3D) anatomical information while reducing reliance on daily CBCT acquisition. We propose a physics-constrained dual-domain network (PCD-Net) for CBCT reconstruction from ultra-sparse orthogonal projections. The framework integrates three key components: (1) a Projection Restoration Network that estimates missing angular information in the projection domain, (2) a differentiable analytic geometry transformation operator that enforces physical consistency between projection and image domains, and (3) a Volumetric Refinement Network that enhances reconstructed CBCT image quality. The method was evaluated on 360 CBCT datasets acquired from 15 GYN cancer patients, with 24 fractions per patient used for patient-specific training and testing. Reconstructed volumes were quantitatively compared with reference CBCT using peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and mean absolute error (MAE) within a body mask. Performance was benchmarked against a conditional generative adversarial network (C-GAN) and a standard volumetric denoising diffusion probabilistic model (DDPM). PCD-Net achieved improved reconstruction performance compared with baseline methods, yielding an average PSNR of 48.02 dB, SSIM of 0.9880, and masked MAE of 24.89 HU. In comparison, the conditional GAN achieved a PSNR of 37.94 dB, SSIM of 0.9546, and MAE of 48.45 HU, while the standard DDPM achieved a PSNR of 38.89 dB, SSIM of 0.9634, and MAE of 46.29 HU. Qualitative evaluation demonstrated reduced reconstruction artifacts and improved anatomical consistency relative to the comparison methods. The proposed PCD-Net enables geometry-consistent volumetric CBCT reconstruction from ultra-sparse orthogonal 2D kV projections. This approach has the potential to reduce imaging dose and treatment time in daily IGRT workflows while preserving clinically meaningful volumetric information, making it particularly suitable for GYN radiotherapy where frequent image guidance is required.

Tags

Topics