Singular Value Decomposition-Based Adaptive Sampling Approximate Message Passing Net for Sparse-View CT Reconstruction

Sparse-viewcomputed tomography (CT) imaging is a promising technique for reducing radiation dose and accelerating data acquisition in medical imaging. However, the challenges of handling a reduced number of projection views persist for both iterative estimation and deep neural reconstruction methods. In this paper, to address these challenges, we present a singular value decomposition-based adaptive sampling approximate message passing network (ASAMP-Net) sparse-view CT imaging method. To achieve multiple sparse views projection within a single scene imaging and alleviate the computational burden, our proposed ASAMP-Net method incorporates an adaptive sampling module into the AMP deep unrolling network. This module dynamically adjusts the data samples used during the learning process, making our method highly adaptable to various projection matrices. Moreover, by decomposing the projection matrix into its principal components, our approach identifies the respective contributions of independent structures. We then select the most significant principal components to construct a projection matrix model with increased orthogonality, thereby enhancing reconstruction performance. Extensive experiments on public datasets demonstrate the superiority of our method. Notably, ASAMP-Net handles various sparse projection views with just a single training process, achieving prominent imaging results compared to other methods in the literature.