SSRNet: A CT Reconstruction Network Based on Sparse Connection and Weight Sharing for Parameters Reduction

In recent years, the neural networks are frequently adopted to address the issues of cone beam CT imaging. However, most of the research so far has been to build neural networks individually either in the image domain or in the projection domain for specific purposes, while the connections between them are not well exploited. A reconstruction network that focuses on the filtering and backward projection process can fully exploit the potential connections between the projection domain and the image domain, and can be plugged into other learned reconstruction models for high-quality results. However, until now most of the existing reconstruction networks have taken the fully-connected layer as the backbone, which leads to an explosive growth in model parameters, especially in reconstruction of cone-beam CT. In this paper, a sparse-sharing-reconstruction network (SSRNet) with sparse connections and multi-group weight sharing is proposed, which can be regarded as a substitute of the filtering and backward projection process and can significantly reduce the number of network parameters up to 0.4% of that of the previous models. The experimental results show that the reconstruction results of the 2D SSRNet are basically consistent with those of the traditional FBP algorithm. The 3D SSRNet outperforms the traditional FDK algorithm at 50 layers away from the midplane, while still keeping the number of network parameters within a manageable range.