Reconstructing 3D voxels from ultra-sparse-view 2D X-ray images using a dual-domain balanced network

Ultra-sparse cone-beam computed tomography (CBCT) reconstruction is important for low-dose and rapid 3D imaging, including C-arm-assisted interventions, yet the Feldkamp-Davis-Kress (FDK) algorithm and simultaneous iterative reconstruction technique (SIRT) degrade under sparse views. We propose a cross-domain framework that reconstructs high-fidelity volumes from as few as 2–4 projections. In the projection domain, a U-Net projection view interpolation module (PVIM) synthesizes missing views to strengthen SIRT constraints; in the voxel domain, a 3D U-Net voxel denoising and refinement module (VDRM) suppresses artifacts and restores fine anatomy. We explicitly analyze load balancing between projection and voxel stages by evaluating different output-view settings and quantifying the accuracy–efficiency trade-off. The method is evaluated on simulated cone-beam projections generated from a public lung computed tomography (CT) dataset, providing a reproducible benchmark for sparse-view reconstruction. Under this simulated evaluation setting, the proposed method consistently outperforms FDK and SIRT in RMSE, SSIM, and PSNR across all evaluated input-view settings. Code available at: https://github.com/MRBXCD/PVIM-VDRM .