Computed tomography (CT) has its irreplaceable function in nondestructive testing and medical diagnosis. In practical CT imaging applications, limited-angle scanning is currently due to X-ray’s potential harm to human but limitation of scanning conditions leading to problem ill-posedness. And there are generally regularization methods and deep learning recommended. This paper presents a novel method combining neural network and truncate singular value decomposition (TSVD) regularization method. Firstly, we acquire a large number of trainable data sets reconstructed by TSVD method, which include some key parameters in limited-angle CT reconstruction problem and regularization parameters. Then, neural network can be used to train the data sets, at the same time, sensitivity analysis of the data that the number of projection angles have the highest sensitivity. Next, to demonstrate the advantages of current novel method, we perform reconstruction using projection data from actual phantom CT scans. We conducted comprehensive comparison between our method and other reconstruction methods. The experimental results indicate that our method further improve quality of reconstructed images and suppress limited-angle artifacts.