Conference Agenda

Overview and details of the sessions of this conference. Please select a date or location to show only sessions at that day or location. Please select a single session for detailed view (with abstracts and downloads if available).

Please note that all times are shown in the time zone of the conference. The current conference time is: 17th Sept 2021, 06:49:26am CEST

Session Overview
Oral Session - Advanced image reconstruction methods 2
Wednesday, 21/July/2021:
2:50pm - 4:30pm

Session Chair: Laurent Desbat
Session Chair: Abhinav Jha
Location: virtual (CEST)

External Resource:
  • Link to the prerecordings
  • Link to the recording of the live session
  • Show help for 'Increase or decrease the abstract text size'
    2:50pm - 3:10pm

    Novel approaches to reconstruction of highly multiplexed data for use in stationary low-dose molecular breast tomosynthesis

    Kjell Erlandsson1, Andrew Wirth2, Ian Baistow2, Kris Thielemans1, Alexander Cherlin2, Brian F Hutton1

    1University College London, United Kingdom; 2Kromek Ltd, County Durham, United Kingdom

    Molecular breast imaging (MBI) can be a useful complement to conventional X-ray mammography. Currently, planar imaging is normally used for this purpose. We are developing a stationary tomosynthesis system for MBI, based on CZT detectors with depth-of-interaction (DOI) capability, and multi-pinhole collimation, which could offer significantly improved contrast compared to planar imaging. A large number of pinholes are used in order to obtain high sensitivity and also improved sampling to compensate for the lack of detector motion. This results in multiplexing (MX), which leads to ambiguity regarding the direction of incidence of the detected g-photons. We have developed various novel approaches to address this problem by performing de-MX either before or during the image reconstruction, aided by the DOI information. We have shown that, by optimising the system geometry, it is possible to gain a factor of 2 in effective sensitivity as compared to a system without MX.

    3:10pm - 3:30pm

    Analytical Covariance Estimation for Iterative CT Reconstruction Methods

    Xiaoyue Guo1,2, Yuxiang Xing1,2, Li Zhang1,2

    1Department of Engineering Physics, Tsinghua University, Beijing, China; 2Key Laboratory of Particle & Radiation Imaging, Tsinghua University, Beijing, China

    Since the distribution of the sinogram is known, we propose a new method to propagate the covariance from projection domain to image domain. For iterative reconstruction, the gradient of objective function must be 0 at the point of convergent solution. Making use of this condition, one can deduce an analytical covariance estimator accordingly. We focused on two typical cases, 1) no penalty term, and 2) non-quadratic penalty term. For case 1, we mainly deal with the ill-condition and solve it by widely used Tikhonov-Phillips method in regularization area. For case 2, we mainly consider a linear approximation for the gradient of non-quadratic penalty, and WLS-TV is analyzed. Both statistical and analytical covariance matrices of a simulated phantom are computed. Analytical estimation of the covariance matrices of a simulated phantom are compared with results estimated from samples. Results indicate that analytical and statistical covariance matrices agree well with each other.

    3:30pm - 3:50pm

    Investigation of Subset Methodologies Applied to Penalised IterativePET Reconstruction

    Robert Twyman1, Simon Arridge2, Kris Thielemans1

    1Institute of Nuclear Medicine, University College London, United Kingdom; 2Department of Computer Science, University College London, United Kingdom

    Subset PET image reconstruction algorithms accelerate reconstruction during early iterations. However, at later iterations, many subset algorithms exhibit limit cycle behaviour leading to undesirable variations between subsequent images, resulting in non-convergence in the absence of step size relaxation. A class of variance reduction algorithms address this issue by incorporating previous subset gradients into the update direction computation. This generates an update direction that is a better approximation of the full objective function direction than standard subset algorithms, while maintaining the same low computational cost. In this work, the impact on reconstruction performance when using a deterministic ordered subset method and two stochastic subset methods is investigated. These subset selection methods are applied to a preconditioned gradient ascent algorithm and three variance reduction algorithms. The ordered subset methodology resulted in superior performance for both subset gradient ascent and two of the variance reduction algorithms during initial passes through the data.

    3:50pm - 4:10pm

    On Stochastic Expectation Maximisation for PET

    Zeljko Kereta1, Robert Twyman2, Simon Arridge1, Kris Thielemans2, Bangti Jin1

    1Computer Science Department, University College London, London, UK; 2Institute of Nuclear Medicine, University College London, London, UK

    Ordered subset variants of statistical iterative recon- struction algorithms for PET can improve the performance in early iterations and thus are popular. However, they suffer from convergence issues, e.g., entering limit cycles. This work consid- ers a stochastic variant of the maximum likelihood expectation maximisation. We adapt the algorithm to PET MAP reconstruc- tion, and for a non spatially separable prior, we combine it with the separable surrogate approach to facilitate the computation of the M-step. Preliminary numerical results indicate that the method is competitive with traditional approaches and enjoys excellent convergence behaviour.

    4:10pm - 4:30pm

    Image Properties Prediction in Nonlinear Model-based Reconstruction using a Perceptron Network

    Wenying Wang, Joseph Webster Stayman, Grace J. Gang

    Johns Hopkins University, United States of America

    Nonlinear reconstruction algorithms have demonstrated superior resolution-to-noise tradeoffs compared to traditional linear reconstruction methods. However, their nonlinear, shift variant, and data-dependent nature complicates performance analysis. Furthermore, there usually lacks a predictive framework for image properties that allows efficient control and optimization of imaging performance. This work quantifies the system response of general nonlinear reconstructions using a quantitative perturbation response metric and develops a data-driven approach for prospective prediction of such properties as a function of varying perturbations (size, shape, and contrast profile), patient anatomy, and algorithmic parameter. The feasibility is demonstrated for a penalized-likelihood reconstruction algorithm with a Huber penalty. We incorporated a compact representation of the imaging system and the perturbation as network input and used a three-layer perceptron network for image property prediction. The predicted perturbation response shows good agreement with empirical measurements. The prediction accuracy is generalizable to all perturbations, anatomical locations, and regularization parameters investigated.

    Contact and Legal Notice · Contact Address:
    Privacy Statement · Conference: Fully3D 2021
    Conference Software - ConfTool Pro 2.6.141
    © 2001 - 2021 by Dr. H. Weinreich, Hamburg, Germany