Learned image reconstruction in X-ray computed tomography

Abstract

The basic idea behind X-ray computed tomography (CT) is, given the change in the intensities of X-ray beams passing through a target object, to reconstruct the image of the object's density that is characterized by the attenuation function. The CT scanning process results in the Radon transform of the attenuation function, therefore the reconstruction problem leads to the inversion of this transform. The inversion can be performed by applying, for example, the filtered backprojection (FBP) formula. However, this classical reconstruction approach often results in images corrupted by artefacts. The learned image reconstruction, based on artificial neural networks, is employed to solve the reconstruction quality issue. This thesis aimed to compare the generalization capabilities of two learned image reconstruction approaches: a two-step approach when FBP-reconstruction is followed by the U-net neural network post-filtering, and a one-step approach when image reconstruction is fully automated with the AUTOMAP deep neural network. The two-step approach was shown to outperform the one-step approach since the latter one requires highly diverse training sets to be able to generalize that poses a problem in the context of CT where the amount of real data available for training is usually limited.