The limited-angle cone-beam Computed Tomography (CT) is often used in C-arm for clinical diagnosis with
the advantages of cheap cost and radiation dose reduction. However, due to incomplete projection data, the 3-dimensional CT images reconstructed by conventional methods, such as the Feldkamp, Davis and Kres (FDK) algorithm , suffer from heavy artifacts and missing features. In this paper, we propose a novel pipeline of neural networks jointly by a FDK-based neural network revisited from W¨urfl et al.’s work  and an image domain U-Net to enhance the 3-dimensional reconstruction quality for limited projection sinogram less than 180 degrees, i.e. 145 degrees in our work. Experimental results, on simulated projections of real-scan CTs, show that the proposed pipeline can reduce some of the major artifacts caused by the limited views while keep the key features, with a 16:60% improvement than W¨urfl et al.’s work on peak signal-to-noise ratio.
The work is partially supported by grant 152 187 0014/IAF311014N and SIPGA.