ADVANCE COMPUTED TO MOGRAPHY FOR BIOMEDICAL IMAGING IN CONVOLUTIONAL NEURAL NETWORKS

Abstract

A new deep convolutional neural network (CNN)-based approach is proposed here for resolving inverse
issues that aren't explicitly stated. These days, the usual method to ill-posed inverse problems is
regularised iterative algorithms. Because of the high computational cost of forward and adjoint
operations and the difficulty in selecting hyper parameters, these approaches are difficult to implement
in reality. When the normal operator (H H, where H is the adjoint of the forward imaging operator, H) of
the forward model is a convolution, unrolled iterative approaches take the form of a CNN (filtering
followed by point-wise non-linearity). As a result of this finding, solving normal-convolutional inverse
issues may be accomplished by the use of direct inversion followed by a CNN. An artifact-free picture
may be obtained by using a combination of multiresolution decomposition and residual learning rather
than straight inversion, which captures the physical model of the system. The suggested network's
performance on parallel beam X-ray computed tomography in synthetic phantoms and genuine
experimental sinograms in sparse-view reconstruction (down to 50 views). Iterative reconstruction of
more realistic phantoms using the proposed network beats total variation-regularized iterative
reconstruction and takes less than one second to complete.