Minimizing the Number of Views in Few-View Computed Tomography: a Deep Learning Approach

In some industrial applications of few-view computed tomography (FVCT), such as baggage inspection at airports, flaw detection in pipelines or the imaging of fast hydrodynamic processes, the image needs to be accurately reconstructed from very small views. The paper continues the authors’ work aimed to minimize the number of views in FVCT. It proposes a method to modernize their earlier algorithm ART-TVS that combines algebraic reconstruction, regularization through total variation norm minimization, and adaptive segmentation. Modernization implies the use of an autoencoder model based on the deep convolutional neural network U-Net for segmentation of interim reconstructions. It is shown in numerical experiment that the new algorithm referred to as ART-TVS-DL (ART-TVS with deep learning based segmentation) accurately reconstructs the 2D Shepp-Logan phantom from 5 views, a QR-code from 12 views, and a pipeline section from only 3 views.