Bio Imaging & Signal Processing Lab
  • BISPL News
  • Professor
  • Members
  • Research
  • Publications
  • Software
    • NIRS-SPM
    • k-t FOCUSS
    • Compressive MUSIC
    • Super-resolution Microscopy
    • Patch Low Rank MRI
    • Sparse SPM
    • ALOHA Inpainting
    • Robust ALOHA
    • ALOHA for MR Recon
    • MR Artifact Removal using Robust ALOHA
    • MR Ghost Artifact correction using ALOHA
    • Geometric GAN
  • Press
  • BISPL Hall of Fame
  • Intranet
  • Seminar
  • Imaging Summit 2017

Latest Research Highlight
​Deep Convolutional Framelets: A General Deep Learning Framework for Inverse Problems 

Ye et al, SIAM Journal on Imaging Sciences (in press), 2018

See more

Congratulations ! Three new papers on deep learning-based image reconstruction will appear in IEEE transactions.

3/28/2018

 
  •  Yoseob Han and Jong Chul Ye,"Framing U-Net via Deep Convolutional Framelets: Application to Sparse-view CT", Special Issue on Machine Learning for Image Reconstruction, IEEE Trans. on Medical Imaging (in press),  2018. ​
  X-ray computed tomography (CT) using sparse projection views is a recent approach to reduce the radiation dose. However, due to the insufficient  projection views, an analytic  reconstruction approach using the filtered back projection (FBP) produces severe streaking artifacts. Recently, deep learning approaches using large receptive field neural networks such as  U-Net have demonstrated impressive performance for sparse-view CT reconstruction. However, theoretical justification is still lacking. 
Inspired by the recent theory of  {\em deep convolutional framelets},  the main goal of this paper is, therefore, to  reveal the 
limitation of U-Net and propose new multi-resolution deep learning schemes. In particular, we show that the alternative U-Net variants such as dual frame and  the tight frame U-Nets satisfy the so-called frame condition which make them better for effective recovery of high frequency edges in sparse view-CT.  Using extensive experiments with  real patient data set, we demonstrate that the new network architectures provide  better reconstruction performance.
Picture
  • Eunhee Kang, Won Chang, Jaejun Yoo, and Jong Chul Ye, "Deep Convolutional Framelet Denosing for Low-Dose CT via Wavelet Residual Network", Special Issue on Machine Learning for Image Reconstruction, IEEE Trans. on Medical Imaging (in press),  2018.
Model based iterative reconstruction (MBIR) algorithms for low-dose X-ray CT are computationally expensive. To address this problem, we recently  proposed a deep convolutional neural network (CNN) for low-dose X-ray CT and won the second place in 2016 AAPM Low-Dose CT Grand Challenge. However, some of the texture were not fully recovered.  To address this problem, here we propose a novel  framelet-based denoising algorithm using wavelet residual network  which synergistically combines the expressive power of deep learning and the performance guarantee from the framelet-based denoising algorithms.  The new algorithms were inspired by the recent interpretation of the deep convolutional neural network (CNN) as a cascaded convolution framelet signal representation. Extensive experimental results confirm that the proposed networks have significantly improved  performance and preserves the detail texture of the original images.
Picture
Picture
  • Dongwook Lee, Jaejun Yoo, Sungho Tak and Jong Chul Ye, "Deep Residual Learning for Accelerated MRI using Magnitude and Phase Networks", IEEE Trans on Biomedical Engineering (in press), Invited Paper for Special Section on Deep Learning, 2018.
Accelerated magnetic resonance (MR) scan acquisition  with compressed sensing (CS) and parallel imaging  is a powerful method to reduce MR imaging scan time. However, many reconstruction algorithms have high computational costs.  To address this, we investigate  deep residual learning networks to remove aliasing artifacts from artifact corrupted images. The deep residual learning networks are composed of magnitude and phase networks  that are separately trained. If both phase and magnitude information are available, the proposed algorithm can work as an iterative  k-space interpolation algorithm using framelet representation.  When  only magnitude data is available, the proposed approach works as an image domain post-processing algorithm. We provide the underlying mathematics to optimize the network structure using recent deep convolutional framelets theory. Comparisons using single and multiple coil show that the proposed residual network provides good reconstruction results with orders of magnitude faster computational time than existing compressed sensing methods.

<<Previous

    BISPL@KAIST

    BISPL is a group of people at KAIST who are eager to dedicate their time and effort to investigate the beauty of bio- and medical imaging with the help of mathematics and physics.

    Our Location

ABOUT US
We are interested in novel bio-imaging and signal processing techniques that can greatly enhance our ability to monitor biological structures and functions. The mathematical tool of BISPL is a new signal processing theory called "compressed sensing". According to compressed sensing theory, perfect reconstruction is possible even from sampling rates dramatically smaller than the Nyquist sampling limit, as long as the non-zero spectral signal is sparse and the samples are obtained with an incoherent basis. From the bio imaging perspective, compressed sensing is a disruptive technology that challenges our current imaging principles. Therefore, our current research activities have been focused on demonstration of new proof-of-concept bioimaging tools for the aforementioned real bio-imaging applications .
CONTACT US
Bio Imaging & Signal Processing Lab. 
Dept. of Bio & Brain Engineering 
KAIST
291 Daehak-ro, Yuseong-gu
Daejeon 305-701, Korea

Copyright (c) 2014, BISPL
All Rights Reserved.

Picture
✕