GCUNET: Combining GNN and CNN for Sinogram Restoration in Low-Dose SPECT Reconstruction

To reduce the potential radiation risk, low-dose Single Photon Emission Computed Tomography (SPECT) is of increasing interest. Many deep learning-based methods have been developed to perform low-dose imaging while maintaining image quality. However, most of the existing methods ignore the unique inner-structure inherent in the original sinogram, limiting their restoration ability. In this paper, we propose a GNN-CNN-UNet (GCUNet) to learn the non-local and local structures of the sinogram using Graph Neural Network (GNN) and Convolutional Neural Network (CNN), respectively, for the task of low-dose SPECT sinogram restoration. In particular, we propose a sinogram-structure-based self-defined neighbors GNN (SSN-GNN) method combined with the Window-KNN-based GNN (W-KNN-GNN) module to construct the underlying graph structure. Afterwards, we employ the maximum likelihood expectation maximization (MLEM) to reconstruct the restored sinogram. The XCAT dataset is used to evaluate the performance of the proposed GCUNet. Experimental results demonstrate that, compared to several reconstruction methods, the proposed method achieves significant improvement in both noise reduction and structure preservation.