X-ray tomography reconstruction with machine learning: a study focusing on accuracy and fidelity

For a long time in computed tomography (CT), noise and missing wedge have been two significant issues prohibiting researchers from obtaining reliable insights into material's intrinsic structures. Though much work has been done to denoise sinograms or recover the missing information, from traditional algorithms to emerging machine learning (ML) methods, most of them focus on perceptual performance, i.e., better visual consistency of data. This metric is adequate for computer vision applications, yet is insufficient for the scientific community where data fidelity is more critical, e.g., in the medical fields. In this work, we are trying to combine ML methods and the inherent properties of sinograms, aiming to achieve both state-of-the-art perceptual performance and high fidelity of the filled data. Distinguished from existing ML architectures, we propose a two-fold model implemented through neural networks: one using generative adversarial networks (GAN) and autoencoder to denoise/inpaint the missing-wedge sinogram, and the other one using convolutional neural networks (CNN) model to enforce the denoised/inpainted sinogram to obey their inherent properties. These two steps may need iterate to achieve consistent results. The results on both simulated and experimental data have demonstrated that our methods have achieved state-of-the-art perceptual performance and high fidelity. Our work further indicates that it is possible to incorporate physics into scientific ML models to make ML models more robust and accurate, significantly benefiting the scientific research aided by ML methods. This work is supported by the LDRD program at the FXI facility at NSLS-II, Brookhaven National Laboratory (BNL).