Advanced Deep Learning Approaches to Enhance Magnetic Resonance Tomography

点击翻译按钮获取中文摘要

With origins in 20th century radio-astronomy, tomography is the basis of modern applications ranging from non-destructive testing, to seismic studies, to medical physics and physiology. These multi-disciplinary applications are not only derived from the same core image science principles, but also represent novel applications of important modern physics discoveries. Magnetic resonance (MR) imaging is an important example that combines developments in nuclear magnetic resonance with tomographic concepts from image science. My previous NSERC research has focussed on improving MR tomography, specifically, on enhancing image quality and throughput. MR imaging is an ideal non-invasive method for imaging biological systems. A fundamental challenge remains the long data acquisition time. To overcome this limitation, I have explored appropriate data undersampling methods. Combining these with advanced image reconstruction methods may allow high-quality images to be obtained rapidly. I have met with much success using advanced approaches like compressed sensing. Since 2014, 32 highly qualified personnel (HQP) have been trained, including three who started faculty positions. In 2018, my group was quick to recognize the role of deep learning (DL) in tomography. Leveraging my substantial image science expertise, I have pioneered advanced DL solutions to reconstruct undersampled MR image data (6 publications). In this next grant cycle, my overall goal is to explore links between acquisition and reconstruction of MR data when using DL methods. My goal is to continue to improve MR image quality and throughput. I will focus on 3 key objectives: 1.Use complex-valued networks and establish appropriately curated raw MR databases. 2.Investigate the role of data sampling strategies on the quality of DL-reconstructed images. 3.Explore the relationship between objective and subjective assessments of image quality. Obj 1 focuses on ensuring that our studies use cutting-edge DL frameworks that support complex numbers and ensuring access to large volumes of data. Obj 2 will critically re-appraise many of the assumptions and simplifications used in current DL studies, such as the use of pseudo-random sampling patterns and constraints on the design of multi-coil imaging arrays. I expect reconsideration of these factors will improve image quality. Finally, Obj 3 will correlate assessments made using objective metrics and subjective (human) assessments, to better understand the relationship between such measures and to find objective measures that better match subjective performance. Over the next 5 years, I will have improved DL networks for MR imaging and expect that this knowledge will have a direct impact on enhancing image quality and throughput. In the longer term and through existing and new collaborations, I will broaden applications to other areas and will continue to train HQP in emerging areas of importance to Canada's academic and commercial sectors.