Iterative Neural Networks for Inverse Problems in Medical Imaging
Il Yong Chun, Member, IEEE
Department of Electrical Engineering, University of Hawai‘i at Mānoa, Honolulu, HI, USA
iychun@hawaii.edu
Abstract
To prevent the spread of COVID-19, chest X-ray computational tomography (CT) became an important tool by reducing false negatives of the widely-used RT-PCR test — a.k.a. the swab test. Population exposure to medical radiation via CT is a significant concern, motivating research on lowering the dose of CT. However, it is challenging to acquire high-quality images in low-dose CT. Duel-energy CT has been increasingly used in many clinical applications such as kidney stone characterization. Dual-energy CT acquires measurements with two different energy spectra, and thus can characterize different constituent material. Achieving accurate material decomposition is critical to maximize the benefit of dual-energy CT.
Iterative neural networks (INN) are rapidly gaining attention for solving inverse problems in imaging, image processing, and computer vision. INNs combine regression NNs and an iterative model-based image reconstruction (MBIR) algorithm that considers imaging physics/image formation and noise statistics in measurements. INNs can apply to a wider range of previously unseen data, compared to non-iterative deep regression NNs; they showed outperforming reconstruction quality over state-of-the-art deep regression NNs and existing MBIR optimization methods. Two INN architectures developed by my group, BCD-Net and Momentum-Net, can improve reconstruction accuracy and/or speed compared to existing INNs, while
guaranteeing convergence under some mild conditions. BCD-Net and Momentum-Net have been successfully applied to diverse imaging problems, including low-dose CT, dual-energy CT, low-count emission tomography,
magnetic resolution imaging, light-field photography (simply put, 4D camera imaging).
This talk will briefly review the BCD-Net and Momentum-Net architectures, explain how they are applied to low-dose CT reconstruction and dual-energy CT decomposition, and introduce their benefits over state-of-the-art deep regression NNs and conventional MBIR methods.
Short Bio
Il Yong Chun is a tenure-track Assistant Professor of Electrical Engineering at the University Hawai’i, Mānoa.
He received B.Eng. degree from Korea University in 2019, and Ph.D. degree from Purdue University in 2015, both in electrical engineering. Prior to joining UHM, he was a Postdoctoral Research Associate in Mathematics, Purdue University, and Research Fellow in Electrical Engineering and Computer Science, The University of Michigan, from 2015 to 2016 and from 2016 to 2019, respectively. During his Ph.D. he worked in Intel Labs, Samsung Advanced Institute of Technology, and Neuroscience Research Institute, as a Research Intern of a Visiting Lecturer. His research interest include machine learning & artificial intelligence, optimization, compressed sensing, and adaptive signal processing, applied to medical imaging, computational photography,
biomedical image computing, and autonomous systems. He has over ten peer-reviewed publications in top-tier journals or conferences and over five invited conference publications.