M3B2  Image Reconstruction Methods I

Wednesday, Nov. 4  10:30-12:30  Pacific Salon 1&2

Session Chair:  Johan Nuyts, KU Leuven, Belgium; Arman Rahmim, Johns Hopkins University, United States

Show/Hide All Abstracts

(10:30) M3B2-1, Simultaneous PET-MRI reconstruction with vectorial second order total generalized variation

F. Knoll1, M. Holler2, T. Koesters1, K. Bredies2, D. K. Sodickson1

1Radiology, NYU School of Medicine, New York, USA
2Department of Mathematics and Scientific Computing, University of Graz, Graz, Austria
While current state of the art PET-MR scanners are able to perform both measurements simultaneously, images are - with the exception of the information from the attenuation map - reconstructed separately and the results are only combined at the visualization stage. We propose a novel iterative joint reconstruction framework that treats MR and PET data on a more equal footing, and that complements the joint data acquisition of current integrated multimodality MR-PET systems. It exploits anatomical correlations between MR and PET images by using a dedicated second order total generalized variation multi-channel regularization functional that couples the two modalities during the image reconstruction step. In particular it uses pointwise Frobenius norm coupling of the gradient and Hessian fields of the two modalities. A primal-dual algorithm is used to obtain a numerical solution. Results are shown for clinical in-vivo FDG PET data that are jointly reconstructed with a simultaneously acquired 3D MPRAGE data set. Sharpness and visibility of fine image features is clearly improved for the PET images with the proposed reconstruction. Strongest improvements can be observed for structures that show distinct MR contrast, like the cortical gray-white matter boundaries. The proposed approach also ensures that features that are only visible in one single modality are not falsely transferred to the other modality. Improvements were lower for the reconstructed MR images, which showed only a slightly improved SNR. The reason is that the clinical MR acquisition protocol does not include acceleration that pushes the conventional reconstruction beyond its capabilities. Preliminary results of our ongoing work indicate that higher MR acceleration rates can be employed when used in the context of a joint multi-modality reconstruction.

(10:45) M3B2-2, MR-Guided Dynamic PET Image Reconstruction with the Kernel Method and Spectral Basis Functions

P. Novosad1, A. J. Reader1,2

1Brain Imaging Center, McGill University, Montreal, Canada
2Division of Imaging Sciences and Biomedical Engineering, St. Thomas' Hospital, King's College London, London, UK
Regularization of iterative reconstruction for fully dynamic PET has often been achieved implicitly by estimating coefficients relating to temporal basis functions, such as dataderived temporal basis functions, wavelet temporal basis functions, or compartmental model based temporal basis functions (direct kinetic parameter estimation). In this work, we propose and evaluate a method for anatomy-guided dynamic PET reconstruction using a joint parameterization of the PET image in terms of spatial basis functions from the kernel method applied to a co-registered MR anatomical image, and temporal basis functions using the spectral analysis method. Since the model of the dynamic image is linear, the EM algorithm can be used to find an estimate for the coefficients. We demonstrate that the proposed method combining both basis functions outperforms reconstruction using either spectral temporal basis functions alone or kernel spatial basis functions alone, offering substantially reduced pixel-level RMSE in post-reconstruction parametric maps. Importantly, some benefits are retained even in the case where structures are present in the emission image but absent in the anatomical image.

(11:00) M3B2-3, An Investigation of Regularization for Basis Image Reconstruction in Spectral CT

B. Chen1, Z. Zhang1, E. Pearson2, E. Sidky1, X. Pan1,3

1Department of Radiology, The University of Chicago, Chicago, USA
2Techna Institute, Toronto, Canada
3Department of Radiation & Cellular Oncology, The University of Chicago, Chicago, USA
Spectral CT adds an additional dimension of energy to the conventional CT imaging. As a result, more than one images are usually to be reconstructed. Conventionally, these images are reconstructed separately as isolated inverse problems, requiring minimum effort in adapting existing reconstruction algorithm. Joint reconstruction, on the other hand, takes advantages of the correlation among the images and seems to be more robust and less demanding on the scanning configuration. In this study, we develop an one-step optimization-based reconstruction method with regularization for the basis images. In a simulation study with a dual kVp scan consisting of two sequential limited-angle acquisition, the results have suggested that the method with the regularization improves the basis images by reducing the crosstalk in the bone regions and rendering more uniform textures in soft tissue regions.

(11:15) M3B2-4, Dual Temporal Regularization Framework Within a 4D PET Image Reconstruction Algorithm for Applications in Oncology

T. Merlin1, D. Visvikis1, F. Lamare2

1INSERM UMR1101, LaTIM, Université de Bretagne Occidentale, Brest, France
2Hôpital de Bordeaux, INCIA, CNRS UMR 5287, Bordeaux, France
Respiratory gating is a useful technique to reduce respiratory motion blur of PET images in oncology, but the images suffer from limited signal to noise ratio (SNR). Previous works proposed the use of temporal regularization between the respiratory-gated frames to tackle this issue. The aim of this study is to extend this approach to dynamic acquisitions by combining, within a 4D reconstruction algorithm, temporal regularization using basis functions applied to both respiratory-gated frames and dynamic acquisition frames. The dynamic dataset was divided into 20 frames, each containing 8 amplitude-based respiratory gates. Different sets of variable temporal basis functions were used, independently handling respiratory motion and tracer dynamics in respiratory-gated 4D dynamic frames, leading to a dual temporal regularization. The proposed reconstruction algorithm simultaneously estimates the set of basis functions at each subset along with their relative coefficients for each image voxel. Quantitative evaluation was performed using dynamic Monte-Carlo simulations of the NCAT phantom including respiratory motion and realistic TACs for the different tissues. The proposed method was compared with the performance of a 3D multi-frame OSEM reconstruction algorithm, and 4D reconstruction incorporating single inter-frame temporal regularization. The proposed method achieved significant improvements in terms of SNR for the respiratory-gated frames, in particular for the early frames of short duration. Results on simulated data achieved standard deviation improvement in the simulated lung lesion of 77% in comparison to the standard 3D reconstruction for similar bias levels. 4D reconstruction incorporating a dual temporal regularization is a promising approach for the reconstruction of oncological dynamic acquisitions. Future work will focus on the evaluation of the proposed methodology on dynamic clinical FDG PET/CT acquisitions of lung cancer patients.

(11:30) M3B2-5, Reconstruction of Pre-List Mode Tomographic Data in PET and SPECT

A. Sitek, G. El Fakhri, H. Sabet

Center for Advanced Medical Imaging Sciences, Massachusetts General Hospital and Harvard Medical School, Boston, MA, USA
The use of statistics in reconstruction of tomographic data became routine in research and clinical applications of positron emission tomography (PET) and single photon emission computed tomography (SPECT). Typically during the course of the reconstruction, detections of counts are modeled as Poisson point processes. The list of detected counts (list-mode data format) is created during the acquisition and later processed and reconstructed using various iterative reconstruction algorithms. In this work we investigate an alternative method of image reconstruction that is applied to unprocessed (raw) outputs from electronic readout and reconstructs images directly from the readout outputs. We refer to unprocessed electronic readout outputs as pre-list mode (PLM) data. Our approach takes to a full extend statistical information available and includes the model of the system response that avoids some of the pitfalls of the standard point spread function (PSF) modeling. The method is tested on computer simulations of the brain PET (bPET) scanner and the feasibility of the proposed reconstruction method is demonstrated using simulations of Jaszczak phantom. In addition to the reconstructed image, the approach provides the estimate of the Bayesian covariance matrix. The PLM data reconstruction is an attractive option that can be used for the statistical data reconstruction with accurate and high-complexity system response modeling.

(11:45) M3B2-6, Motion Compensation and Pose Measurement Uncertainty in Awake Small Animal Positron Emission Tomography Using Stochastic Origin Ensembles

J. E. Gillam1, G. I. Angelis1, R. Fulton1,2,3, A. Kyme1,4, S. R. Meikle1

1Faculty of Health Sciences (BMRI), The University of Sydney, Camperdown, NSW, Australia
2School of Physics, The University of Sydney, Darlington, NSW, Australia
3Department of Medical Physics, Westmead Hospital, Westmead, NSW, Australia
4Biomedical Engineering, UC Davis, Davis, CA, USA
In order to remove the influence of anaesthetic agents on neurological function and to expand the range of imaging tasks available it is preferable to image small animals in an awake state. Accurate activity estimates in awake small animal Positron Emission Tomography rely on the measurement of animal head pose over the time-course of the scan. Pose measurements are then incorporated into the emission data during image reconstruction, compensating for animal motion. Uncertainty in pose measurement can impact reconstructed image quality by effectively degrading data resolution, and hence that of the reconstructed image. In small animal imaging, regions of interest can be small in comparison to image-space voxelisation so that the precision of estimates taken from the reconstructed image can be difficult to gauge. Stochastic Origin Ensembles provides a means of estimating a more complete statistical description of the emission data than other methods of image reconstruction. In this investigation, rigid motion compensation is incorporated into the Stochastic Origin Ensembles algorithm and explored using simulated data. Realistic motion is modeled within a GATE simulation and measurement uncertainty is incorporated into the pose data using both simulated perturbations as well as experimental trials using a motion tracking system. Sampling from the posterior distribution is conducted using the Stochastic Origin Ensembles algorithm and compared to image reconstruction using the Maximum Likelihood-Expectation Maximisation algorithm. Using Stochastic Origin Ensembles both regional and single voxel parameters are investigated. The impact of varying levels of pose measurement uncertainty on image-space parameters is demonstrated and assessed.

(12:00) M3B2-7, Reconstruction of CT Images from Sparse-View Polyenergetic Data Using Total Variation Minimization

T. D. Humphries, A. Faridani

Mathematics, Oregon State University, Corvallis, OR, USA
Recent work in CT image reconstruction has seen increasing interest in the use of compressive sensing techniques to reconstruct images from sparse-view projection data, with the goal of reducing radiation dose as well as scan time. Most often these reconstruction approaches exploit sparsity in the gradient of the image using total variation (TV) minimization. Following the existing theoretical results from compressive sensing, these approaches typically assume a linear measurement model, which corresponds to data generated from a monoenergetic X-ray beam. Most clinical CT systems generate X-rays from a polyenergetic spectrum, however, which is inconsistent with a linear system model and produces the well-known beam hardening artifacts. Such artifacts have been observed in some studies on sparse-view CT reconstruction using a linear model. In this work we incorporate an existing polyenergetic iterative technique known as polyenergetic SART (pSART) into a TV minimization reconstruction algorithm. Using numerical phantom experiments, we demonstrate that this polyenergetic TV minimization algorithm is able to reconstruct images free of both undersampling and beam hardening artifacts from sparse-view, polyenergetic projection data.

(12:15) M3B2-8, Low-Dose CT Image Reconstruction Method With Probabilistic Atlas Prior

M. Selim1, H. Kudo2, E. A. Rashed3

1Dept. of Mathematics and Computer Science, Faculty of Science, Suez University, Suez, Egypt
2Division of Information Engineering, Faculty of Engineering, Information and Systems, University of Tsukuba, Tsukuba, Japan
3Image Science Lab., Dept. of Mathematics, Faculty of Science, Suez Canal University, Ismailia, Egypt
This work investigates the problem of image reconstruction from low-dose x-ray computed tomography (CT). Statistical iterative reconstruction is known to provide higher image quality due to the ability to incorporate prior knowledge to the reconstruction method and accurately model the photon statistics. In this paper, we develop a statistical reconstruction method using prior knowledge extracted from probabilistic atlas. First, we use a set of CT images previously scanned of various patients to generate a probabilistic atlas using Gaussian mixture model (GMM). Then, expectation maximization (EM) clustering algorithm is used to estimate the mixture parameters. Probabilistic atlas and mixture model parameters are then used to formulate the image reconstruction cost function. By merging the atlas information and smoothing penalty into the reconstruction procedure, image quality has been remarkably improved.