Image Reconstruction with Variational Networks: Application to Synchrotron Radiation Imaging

摘要

In computed tomography, image reconstructed from limited projection data is subject to strong noise and artifacts. Reducing the number of projection views in synchrotron radiation imaging has a great benefits in decreasing the computation cost for image formation. Moreover, it also prevents the damage of biological specimen caused by the x-ray radiation exposure. In this paper, a new procedure for image reconstruction from limited projection views with the assist of deep learning approach is proposed. A deep learning approach is employed to estimate an initial guess of the tomographic image followed by a fast row-action reconstruction to guarantee the restoration of any pattern missed in the first stage. Most of previous attempts that use deep learning in image reconstruction are implemented as a post-processing techniques applied on the reconstructed image. This may lead to anomalies in the final result due to the lack of consistency between the acquired projection data and the reconstructed image. The proposed procedure solves this problem by keeping the data consistency up to the final stage of image formation. This would reduce the possibility of losing image abnormal structures due to insufficient training in deep learning. Experimental results using synchrotron radiation data demonstrate the usefulness of the proposed framework.