Sparse-View Cone-Beam CT Image Reconstruction Based on Conditional Generative Adversarial Networks for Sinogram Restoration

Sparse - view projection in cone - beam CT (CBCT) cuts dose and time but causes streak artifacts. Sinogram inpainting can generate missing data for better image quality. A residual encoder-decoder generative adversarial network (Pix2pixGAN) is proposed for sparse - view CBCT reconstruction. It replaces the U - Net generator with a residual encoder - decoder module and uses a PatchGAN - based discriminator. Trained with real CBCT data, Pix2pixGAN was tested at 1/2, 1/3, and 1/4 sparse sampling, compared with linear interpolation and the original Pix2pixGAN. Results show it performs best, especially at 1/4 sampling. In the sinogram domain, RMSE reduces by 7.2%, PSNR increases by 1.5%, and SSIM improves by 1.4%. In the reconstruction image domain, RMSE drops by 5.4%, PSNR rises by 1.6%, and SSIM improves by 1.0%. In conclusion, Pix2pixGAN is suitable for high - quality sparse - angle CBCT reconstruction and has potential for fast, low - dose scanning.