Deep Learning Based CT Image Reconstruction

Thesis (Master's)--University of Washington, 2020As a common medical imaging method, Computed Tomography (CT) can create tomographic images using X-ray data acquired from around the human body. However, high quality and adequately sampled X-ray measurement data are not always available. In this scenario, the tomographic image created by conventional reconstruction algorithms will be noisy, or contain artifacts. The goal of our study is to reconstruct high-quality tomographic images from noisy or incomplete scan data, including low-dose, sparse-view, and limited-angle scenarios, by utilizing novel deep learning techniques. In this project, we trained a Generative Adversarial Network (GAN) and used it as a signal prior in the Simultaneous Algebraic Reconstruction Technique (SART) iterative reconstruction algorithm. The GAN we trained includes a self-attention block to model long-range dependencies in the scan data. Compared with the state-of-the-art denoising cycle GAN, CIRCLE GAN, and a conventional mathematical reconstruction algorithm incorporating total variation minimization, our Self-Attention GAN for CT image reconstruction produces competitive results on solving limited-angle data reconstruction problems. On sparse view and low-dose scenarios, our method is not always the best among compared methods, but produces competitive results in some situations