Three-dimensional residual projection completion for coded-aperture cone-beam computed tomography

Coded aperture cone-beam computed tomography reduces X-ray exposure by blocking a subset of detector-view rays, but the resulting interleaved missing samples make direct volumetric reconstruction computationally demanding and prone to artifacts. This paper presents a projection-domain reconstruction framework in which a 3D U-Net learns residual corrections to complete the coded projection tensor before Feldkamp-Davis-Kress reconstruction. The network takes the measured coded projection, an interpolation-based base estimate, and the binary aperture mask as input, and it predicts a bounded residual correction only for unmeasured samples while preserving measured data exactly. A training objective combining missing-region fidelity, seam intensity consistency, and seam gradient consistency is used to suppress discontinuities around coded-aperture boundaries. Experiments on 188 projection volumes show that the selected loss composition improves projection completion and yields reconstruction results close to the reference FDK images. Over 38 testing cases, the mean Ref-vs-Pred reconstruction PSNR is 44.6 dB and the mean SSIM value is 0.98.