Date of Award
Spring 2024
Document Type
Dissertation
Degree Name
Doctor of Philosophy (PhD)
Department
Biomedical Engineering (ENAS)
First Advisor
Liu, Chi
Abstract
PET-CT/MRI is the most used multi-modal tomography imaging exam which can provide both functional and anatomical information from a single exam and aid clinical decision-making. The CT/MRI component focuses on visualizing the anatomical structures, and the PET component focuses on visualizing molecular-level functional activities in tissues. Millions of PET-CT/MRI scans were performed each year worldwide with wide applications in oncology, cardiology, neurology, and biomedical research. While being an indispensable multi-modal tomography imaging tool in medicine, PET-CT/MRI has several key issues limiting its broader impacts on patients, including 1) the potential hazard caused by radiation dose from the PET and CT, 2) the degraded image quality caused by reduced dose, metal implants, and motions, and 3) the prolonged acquisition time with increased motion and patient discomfort. Therefore, this dissertation aims to address these challenges by developing a line of deep-learning techniques for PET-CT/MRI radiation dose, image artifacts, and acquisition time reductions. Starting with CT, we first proposed a cascade reconstruction network with projection data fidelity for CT acquired with a reduced number of X-ray projections, thus reducing the radiation dose of the CT component. To address the metal artifacts under the low-dose acquisition conditions, we built up the concepts of dual-domain learning that learn signal restoration in both the image domain and the original data acquisition domain, i.e. sinogram. In PET imaging, to reduce the radiation dose and motion, we proposed the first AI reconstruction framework for low-dose gated PET imaging. We devised a unified motion correction and denoising deep network that allows joint optimization of motion estimation/correction among low-dose gated images and denoising of the motion-compensated image for high-quality PET reconstruction. To further reduce the PET acquisition time and correct motion regardless of type, we developed a deep-learning-aided reconstruction framework that allows modeling-free quasi-continuous motion estimation via a deep registration model, and conversion from short to long-acquisition image via a deep generative model. To enable training from multi-institutional data for obtaining a robust deep denoising model for PET, we also devised the first personalized deep denoising solution for multi-institutional co-training with no data sharing. To further reduce radiation in PET/CT, we built a population-prior-aided deep generation method for generating the attenuation map directly from low-dose PET to eliminate the need for CT for PET attenuation correction. Lastly, to accelerate MRI acquisition, we developed a dual-domain self-supervised learning scheme that allows high-quality accelerated MRI reconstruction without fully-sample k-space data as ground truth. In summary, the proposed techniques each aim to address a specific set of challenges in PET-CT/MRI, collectively adding new insights into how we can use AI to transform nuclear medicine imaging into a more safe, efficient, and high-quality exam tool for patient healthcare.
Recommended Citation
Zhou, Bo, "Deep Learning for Multi-Modal Tomography Imaging: Radiation Dose, Image Artifacts, and Acquisition Time Reductions" (2024). Yale Graduate School of Arts and Sciences Dissertations. 1476.
https://elischolar.library.yale.edu/gsas_dissertations/1476
