+

CSE Medical Imaging Seminar

Deep learning based methods for artifact correction in X-ray CT

Hyoung Suk Park / Senior Researcher, National Institute for Mathematical Sciences (NIMS)

Organizer Jin Keun Seo seoj@yonsei.ac.kr

Contact Haeeun Han csedept@yonsei.ac.kr

Friday 4th, 11th, 18th May 2018 ASTC Room 615, Yonsei University 연세대학교 첨단관(117동) 615호

2018 CSE MI Seminar Series

Friday, May 04, 2018, at 11:00

1

Recently, deep learning techniques show significant performance over existing methods for medical imaging modalities including X-ray

computed tomography (CT). In this talk, I will provide some overview of deep learning algorithms including convolutional neural network. I

will then introduce how deep learning is utilized for addressing problems in X-ray CT imaging (e.g., artifacts arising from low dose or metallic

objects). I also introduce about our deep learning approach for metal artifact reduction. Taking account of the fundamental difficulty in obtain-

ing sufficient training data in a medical environment, the proposed learning method is designed to use simulated training data and a

patient-type specific learning model is used to simplify the learning process. We show the feasibility of the proposed method using numerical

simulations and phantom experiment.

Medical imaging is the technique and process of visualizing the anatomy of a body for clinical analysis and medical intervention, as well

as the function of some organs and tissues. However, the reconstructed image in general suffers from the severe artifacts due to the ill

posed nature of underlying linear inverse problem. In this talk, I will briefly introduce some related topics in the inverse problem in medical

imaging based on my works. First, I will give the mathematical analysis of the inverse problem in quantitative susceptibility mapping. In

the following, I will present "how to solve inverse problem" in the following two aspects. One is the edge driven wavelet frame based image

restoration model which is designed to restore/enhance the key features in a given image. Finally, the harmonic incompatibility removal

(HIRE) susceptibility reconstruction model will be presented.

With the introduction of X-ray flat panel detectors, cone-beam computed tomography (CBCT) has been widely used and accelerated the

utilization of three-dimensional (3-D) images for clinical diagnosis. The diagnostic accuracy is closely related to the imaging performance,

and thus the optimization of imaging systems and image processing algorithms has been an important issue. For linear reconstruction (e.g.,

FDK), image quality or imaging performance can be evaluated by Fourier based metrics (MTF, NPS, and NEQ), and these metrics can be utilized

for a task based image quality assessment. In this talk, I will cover three sub topics : 1) A method to measure 3D MTF of a CBCT system, 2)

NPS measurement of CT systems(i.e., parallel, fan, and cone beam CT) and 3) mathematical model observer and its applications on image

quality assessment of CT images.

+

Image Quality Assessment for Computed Tomography Images

Jongduk Baek / Associate Professor, School of Integrated Technology, Yonsei University

Friday, May 11, 2018, at 13:30

+

Inverse Problem in Medical Imaging and Beyond

Jae Kyu Choi / Postdoc Researcher, Shanghai Jiao Tong University

Friday, May 11, 2018, at 10:30

+

Accelerated Gradient Methods for Large-scale Convex Optimization

Donghwan Kim / Research Instructor, Department of Mathematics, Dartmouth College

Friday, May 18, 2018, at 11:00

Many modern applications such as machine learning, inverse problems, and control require solving large-dimensional optimization prob-

lems. First-order methods such as a gradient method are widely used to solve such large-scale problems, since their computational cost per

iteration mildly depends on the problem dimension. However, they suffer from slow convergence rates, compared to second-order methods

such as Newton's method. Therefore, accelerating a gradient method has received a great interest in the optimization community, and this

led to development and extension of a conjugate gradient method, a heavy-ball method, and Nesterov's fast gradient method, which we

review in this talk. This talk will then present new proposed accelerated gradient methods, named optimized gradient method (OGM) and

OGM-G, that have the best known worst-case convergence rates for smooth convex optimization among any accelerated gradient methods.