An improving reconstruction network for sparse-view spectral CT based on dual-domain hybrid encoder integrating ResNet and Transformer

Sparse-view Computed Tomography (CT) is an effective method for reducing radiation dose. However, the incompleteness of the projection results in streaking artifacts in the reconstructed images. Dual energy CT Data complementarity and diversity in both the projection and image domains represent a promising solution to recover the missing information from sparse views. We proposed an end-to-end sparse-view dual-domain improved dual-energy CT reconstruction network (DdDe-Net) to recover the missing information due to the reduced number of projections. The network is comprised of a projection domain data correction module (PDDC-Net) and an image domain data correction module (IDDC-Net). In PDDC-Net, ResNet and Transformer architectures are combined with a residual module to create a hybrid encoder structure, effectively suppressing noise and artifacts in the projection domain. The IDDC-Net further reduces streaking artifacts and preserves tissue details through the prior image and differential learning process. DdDe-Net was trained and validated on a sparse-view extended dataset of simulated breast DECT data and an abdominal slice simulation dataset. Compared to several mainstream CT image reconstruction algorithms, our proposed algorithm demonstrates a notable advantage, achieving a RMSE of 0.0118 ± 0.0032, a SSIM of 0.9688 ± 0.0177, and a PSNR of 39.83 ± 1.14 dB. The experimental results indicate that DdDe-Net effectively mitigates artifacts while preserving image details in sparse-view DECT, underscoring its potential for practical applications.