Strategies of Deep Learning for Tomographic Reconstruction

Abstract

In this article, we introduce three different strategies of tomographic reconstruction based on deep learning. These algorithms are model-based learning for iterative optimization. We discuss the basic principles of developing these algorithms. The performance of them is analyzed and evaluated both on theory and simulation reconstruction. We developed open-source software to run these algorithms in the same framework. From the simulation results, all these deep learning algorithms showed improvements in reconstruction quality and accuracy where the strategy based on Generative Adversarial Networks showed the advantage especially.

Physics-informed generative network for phase retrieval from a single intensity measurement

“…GANrec significantly reduced the chance of getting stuck in local minima by using the discriminator rather than a simple loss function [48]. We observed a good quality of the phase retrieval from the evaluations of both simulations and real measurement data.…”

Calibration by differentiation – Self‐supervised calibration for X‐ray microscopy using a differentiable cone‐beam reconstruction operator

“…The embedding of knowledge about the physics underlying CT reconstruction into learned pipelines has been explored before, mostly in the medical context, but also on XRM data similar to ours. [32][33][34] The areas of application range from improving reconstruction quality in low dose, limited or missing angle scenarios 34,35 over learning of data-optimal reconstruction filters or weights 36,37 to calibration of, for example, the position of the rotation axis from the measured data. 32 Some of the mentioned approaches propose deep learning architectures in the classical sense involving a high number of trainable parameters.…”

A scalable neural network architecture for self-supervised tomographic image reconstruction

“…[7][8][9] A number of notable exceptions do exist, where supervised learning, and generative models have been used to automatically map from sinogram to real space. 5,[10][11][12][13][14][15][16] While these methods are very promising, bottlenecks still exist to their application to image reconstruction due to their scalability (i.e. their ability to handle large images), their network size (large networks can be computationally very expensive) and particularly for applications where absolute values (as opposed to normalised values) are important in the reconstructed image, such as in chemical tomography and in quantitative analysis of attenuation-based tomography data.…”