IEEE 2017 NSS/MIC/RTSD ControlCenter

Online Program Overview Session: M-18

To search for a specific ID please enter the hash sign followed by the ID number (e.g. #123).

Param. Imaging and Kinetic Modeling

Session chair: Richard E. Carson Yale University; Quanzheng Li Harvard University
Shortcut: M-18
Date: Saturday, October 28, 2017, 08:00
Room: Centennial III
Session type: MIC Session


8:00 am M-18-1

Strategies to improve Direct EM Patlak reconstructions (#2473)

J. - D. Gallezot1, Y. Lu1, K. Fontaine1, C. Liu1, R. E. Carson1

1 Yale University, Department of Radiology and Biomedical Imaging / Yale PET Center, New Haven, Connecticut, United States of America


Direct Expectation-Maximization (EM) reconstruction of parametric images from PET data has been proposed, to take advantage of the known Poisson character of raw PET data. However, direct reconstruction algorithms are then dependent on the properties of the kinetic model and of the radiotracer used. Moreover, one of the first direct EM reconstruction algorithm proposed, the Parametric Iterative Reconstruction (PIR) algorithm, based on the Patlak plot, converges slowly when applied to [18F]FDG data. Thus, some alternative algorithms have been proposed to accelerate convergence. In this study, we propose additional strategies to improve the PIR algorithm. One strategy is to use a line search procedure to increase the size of steps taken at each iteration (PIR-LS). Another group of strategies is to apply physiological constraints (PIR-PC) on the intercept parameter of the Patlak plot, since this parameter has a small dynamic range and is more stable across subjects and across physiological and disease states. These constraints could be of two types: 1) only the Ki parameter is updated during the first iterations; 2) penalty terms are added to the log likelihood function. The prior value for the intercept parameter can be assigned to each voxel based on the CT image (PIR-PC-CT), or on a coregistered Atlas (PIR-PC-A), or on the ratio of tissue to input function values (PIR-PC-R) just before the start of the linear part of the Patlak plot. PIR-LS and PIR-PC were first compared to PIR and the Nested EM algorithm on simulated 2D data. Results showed that both methods could increase the convergence speed of PIR. Then, PIR and PIR-LS were compared using a simulated dynamic brain dataset. PIR-LS images had higher contrast between white and grey matter at the same number of iterations. Finally, PIR and PIR-PC-R were compared on a real lung tumor data set. With PIR-PC-R, Ki values were higher in the tumor at the same number of iterations, consistent with faster convergence.

Keywords: Patlak, Parametric images, Direct reconstruction, Expectation-Maximization, FDG
8:18 am M-18-2

Cluster-based Direct Estimation of Parametric Maps of Dopamine Response in Dynamic PET Data (#2760)

G. I. Angelis1, S. R. Meikle1

1 The University of Sydney, Brain and Mind Centre, Faculty of Health Sciences, Sydney, New South Wales, Australia


Advanced kinetic models, such as the linear parametric neurotransmitter PET (lp-ntPET), have been recently developed to model and quantify transient changes in the radiotracer efflux from the target tissue, caused by endogenous neurotransmitter release during an cognitive task or environmental stimulus. Nevertheless, application of these models to voxel-wise time-activity curves is often challenging due to high levels of noise, which substantially increases the likelihood of generating false positive findings. In this paper we propose and a direct reconstruction framework which includes a cluster-based model selection before the parameter fitting step to allow the use of the lp-ntPET model only to those voxels with significant activation response. We use principal component analysis and Gaussian mixture models to determine clusters of voxels which share similar properties (e.g. similar activation response), with the ultimate aim to reduce the false positive responses in brain regions other than the target (e.g. striatum). We used computer simulations based on a digital rat brain phantom to generate dynamic PET data, representing a [11C]raclopride study, with a known activation response whose magnitude was similar to an endogenous dopamine release (50% increase over the baseline). Results showed that the proposed direct reconstruction framework can define the activated voxels with high accuracy, minimising false positive responses outside the target region, compared to its post reconstruction counterpart. In addition, it led to more accurate voxel-wise activation profiles, with clear temporal information. Work is currently in progress to apply this methodology to real rat data of unknown activation response.

Keywords: Direct parameter estimation, neurotransmitter response, lp-ntPET, non-equilibrium kinetics, 4D reconstruction
8:36 am M-18-3

SUV/Patlak-4D whole-body PET/CT dynamic and parametric imaging: clinical demonstration and validation of SUV synthesis from dynamic passes (#3894)

N. A. Karakatsanis1, M. E. Casey2, K. Knesaurek3, Z. A. Fayad1, L. Kostakoglu3

1 Icahn School of Medicine at Mount Sinai, Translational and Molecular Imaging Institute, New York, New York, United States of America
2 Siemens Healthineers, Division of Molecular Imaging, Knoxville, Tennessee, United States of America
3 Icahn School of Medicine at Mount Sinai, Division of Nuclear Medicine, Department of Radiology, New York, New York, United States of America


Whole-body (WB) PET/CT imaging currently involves single-frame, i.e. static, PET acquisitions over a set of axial bed positions to form standardized uptake value (SUV) images. However, SUV is considered semi-quantitative due to its dependence on post-injection scan time and subject physiology. Recently, we introduced a class of novel WB dynamic PET/CT scan protocols capable of acquiring PET data over multiple time frames at each bed to enable quantitative WB parametric PET imaging. In this study, we are validating a special dynamic WB 18F-FDG PET/CT framework supporting synthesis of static-equivalent SUV as well as dynamic and parametric Patlak WB images from a single set of WB passes acquired within the standard-of-care post-injection time window of 18F-FDG PET/CT exams. The proposed protocol, streamlined for wide clinical adoption, consists of four unidirectional WB passes of constant bed frames (30 sec) or equivalent bed speeds, beginning 1h after injection. The acquired data at each bed may later be used by a direct 4D Patlak reconstruction algorithm to obtain parametric WB Ki images after utilizing a population-based model to infer the missing first 1h of the input function. Moreover, the same data can be added at each bed to synthesize static PET frames and equivalent SUV WB PET images. For the clinical evaluation of the proposed method, we consented patients scheduled for their regular WB 18F-FDG PET/CT exam to undergo afterwards a second PET-only scan, for the same total scan time, using our protocol. SUV mean comparisons at focal uptake ROIs between the step-and-shoot (S&S) static SUV and the respective synthetic SUV images yielded <14% and <20%.differences, primarily attributed to post-smoothing, when synthesis was conducted on S&S and continuous bed motion scan modes respectively. In addition, the direct 4D Patlak Ki WB images exhibited similar noise levels, while offering similar or better lesion contrast and enhanced quantification relative to SUV.

Keywords: dynamic, whole-body, PET, PET/CT, parametric, SUV, Patlak, direct, 4D
8:54 am M-18-4

Joint Optimization of Kinetic Modelling and CBM Acquisition Parameters in Hybrid Whole-Body Dynamic PET Imaging (#4080)

F. A. Kotasidis1, M. Manari1, V. Garibotto1, H. Zaidi1

1 Geneva University Hospital, Division of Nuclear Medicine & Molecular Imaging, Geneva, Switzerland


In our previous work we proposed a modilfied protocol for simultaneous estimation of compartmental and graphical analysis kinetic parameters in whole-body (WB) dynamic PET imaging. The protocol is clinically feasible in continuous bed motion (CBM) acquisition mode. Therefore certain kinetic modelling related parameters, such as temporal frame definition and duration are selected a priori through CBM pass and bed speed selection. This is in contrast to traditional single bed dynamic imaging where such parameters are defined a posteriori. Furthermore optimum CBM parameters might vary between compartmental and graphical analysis. To this end a joint optimization of acquisition and kinetic modelling parameters is attempted. Using the XCAT phantom as well as clinical studies, the impact of temporal frame definition, frame duration and total scan duration on micro- and macro- parametric images is evaluated, through modulation of CBM parameters such as the bed speed and number of passes. Furthermore for Patlak graphical analysis and compartmental modelling, the optimum regression window was systematically investigated with respect to the selected CBM parameters.

Keywords: kinetic modelling, wholdbody imaging, parametric imaging, continuous bed motion
9:12 am M-18-5

Robust Nonlinear Parameter Estimation in Tracer Kinetic Analysis via Infinity Norm Regularization (#2267)

S. K. Kang1, S. Seo2, J. S. Lee1

1 Seoul National University, Seoul, Republic of Korea
2 Gachon University, Incheon, Republic of Korea


The generation of parametric images of individual rate constants describing the tracer kinetics is challenging because the noise level of voxel time-activity curve is very high in PET. Moreover, the complex and nonlinear relationship between the rate constants and PET time-activity curves usually leads to the biased and dispersed estimation results. In this study, we have identified that the distribution of estimated parameters in PET kinetic models is skewed under highly-noisy circumstances. Therefore, we’d like to propose an effective mitigation method for this problem via infinity norm regularization. Conventional optimization method such as Levenberg-Marquardt Algorithm (LMA) is inappropriate to minimize the given non-smooth cost function. Therefore, we examined two estimation methods that do not require smoothness of cost function to be combined with the infinity norm regularization. One of them is particle swarm optimization (PSO), and the other is proximal gradient algorithm. 

A simulation study was performed for the three compartment model with four kinetic parameters. To explore the effects of noise levels to those algorithms, noise with various levels was added to noiseless time-activity curves and kinetic parameters estimated using different estimation algorithms.

The proximal gradient could not reach the ground truth even with the fairly good initial parameters. Although PSO’s initial values were randomly selected, it searched the solution much better than the proximal operator. The proximal gradient algorithms were extremely sensitive to the initial value, and tuning inner parameter (e.g. step size) was very difficult. It would be because our problem is highly nonlinear particularly in the high noise level. The PSO showed better results than single and repeated LMA. In particular, the regularized PSO showed the best performance at all noise levels.

Keywords: Tracer kinetic analysis, infinity norm regularization, particle swam optimization
9:30 am M-18-6

Cardiac Phantom for Improved Small-Animal Dynamic SPECT Myocardial Blood Flow Quantification (#2988)

L. C. Johnson1, M. Guerraty2, D. Matej1, S. D. Metzler1

1 University of Pennsylvania, Department of Radiology, Philadelphia, Pennsylvania, United States of America
2 University of Pennsylvania, Division of Cardiovascular Medicine, Philadelphia, Pennsylvania, United States of America


To assess myocardial blood flow (MBF) in mice, and subsequently coronary microvascular function, we have previously worked to develop a non-invasive, in-vivo, dynamic cardiac imaging protocol using the U-SPECT+ system. This work investigates microSPECT quantitative capabilities of cardiac imaging by utilizing a multi-part cardiac phantom and applying a known kinetic model to static data, allowing for assessment of kinetic modeling accuracy. A multi-part cardiac phantom was designed with four main components: two left-ventricular myocardium sections and two left-ventricular blood pool sections, sized for end-systole and end-diastole. Each section of the phantom was imaged separately while acquiring list-mode data. A Gaussian-based kinetic model was developed and applied to list-mode data to estimate both the myocardium and blood pool activity concentrations over time. K1 and the injection time length were chosen to match that of previous animal data. The generated dynamic list-mode data were then combined, such that the myocardium and blood pool segments are contained in a single image, and reconstructed. Regions of interest were drawn in the resulting images, and kinetic modeling parameter estimation was performed. Given an input K1 of 0.005 (mL/sec/g), the resulting fit estimated K1 of 0.0043 and 0.00533 mL/sec/g for end-diastole and end-systole, respectively. We have designed and imaged a phantom that allows for estimation of the of accuracy of kinetic modeling of small animal cardiac SPECT and performed an initial evaluation. Future work will employ multiple reconstruction noise ensembles to analyze the robustness of our fitting method and to evaluate the effects of noise.

Keywords: Kinetic Modeling, SPECT, Small-Animal, Phantom, Quantification Accuracy, Myocardial Blood Flow Reserve, Myocardial Blood Flow