PET transmission tomography using a novel nonlocal MRF prior☆
Introduction
In positron emission tomography (PET), a chemical compound labeled with a positron emitting radioisotope is injected into the bloodstream of the patient. The purpose is to obtain an image of the concentration of that chemical compound which is related to a biological function in the patient’s body. The radiotracer nucleus change from a metastable state to a stable state by emiting a positron (positively charged electron). The emitted positron annihilates with a nearby electron to form a pair of 511 keV photons propagating in opposite directions. A positron-electron annihilation is considered to take place in the line joining two detectors when such a pair of photons are detected almost simultaneously outside the subject. This simultaneous detection is termed coincidence detection. And each coincidence detection increments a counter that represents the line integral specified by two detectors. After a certain scan time, line integral detected sinogram data of the radioisotope density are obtained [1], [2].
However, not all annihilations that result in a photon pair heading towards two detectors are detected. Often, one or both of the photons get scattered or absorbed by the patient body resulting in no detection. The survival probability of an annihilation event is determined by the length and the type of the tissue that the photons traverse. This effect is called attenuation. The attenuation map or attenuation correction factors (ACFs) express spatially the mass absorption coefficients for the transaxial slice of the body. Such attenuation is different for different tissue types, hence the measurements should be compensated for attenuation for each ray (or line integral) [3].
Conventionally, the ACFs are computed from the blank and the transmission sinograms by dividing them. In clinical practice, short scan times with precise attenuation correction are requested for quantitative whole body studies because statistically desired long acquisition times for transmission are clinically impractical and inconvenient to patients. However, short scan times with low total counts levels often bring statistical noise to reconstructions. So, the short scan data are conventionally smoothed with the resulting unfavorable blurring of details propagating to the emission sinogram [3], [4], [5].
Attenuation correction can also be performed by computational methods [3], [6]. The attenuating matter is approximated by an area with a uniform value of linear mass absorption coefficient . The ACFs can be computed by projecting the graphically defined -image and taking the exponential:
For above computational methods, when filtered back projection (FBP) is used in count-limited reconstructions, noise and some bias may be introduced in the image [6]. In order to remove the noise from the transmission image, the segmented attenuation correction method classifies pixels of the image to certain tissue types with typical attenuation values before the projection. But the parameter-relevant segmentations are rather complicated and might be unpredictable especially for the noisy FBP reconstruction. In practice, some details may not properly contribute to the ACFs and noise on the other hand may wrongly contribute to the ACFs especially in the case of short scan [7], [8], [9].
Bayesian methods, which have been widely accepted for producing better reconstructions, are also recommended for transmission tomography [10], [11], [12]. Ollinger proposed one step of Newton’s method for transmission reconstruction [9]. And Lange put forward an one step late expectation-maximization (OSL-EM) iterative reconstruction algorithm for Bayesian reconstruction of transmission tomography [10]. In the past 20 years, Fessler and his group have done a great deal of remarkable work on emission and transmission tompgraphies. They have devised effective convergent iterative algorithms for PET transmission tomography [11], [12], [13], [14].
As to the Bayesian reconstruction for transmission tomography, the simple quadratic membrane (QM) smoothing prior, which smoothes both noise and edge details equally, tends to produce an unfavorable oversmoothing effect. Considering the fact that the transmission attenuation maps are often composed of homogeneous regions with sharp boundaries, edge-preserving nonquadratic priors, which are able to produce sharp edges by choosing a nonquadratic prior energy, have been widely used in the studies of CT or PET transmission tomography [10], [11], [12], [13], [14], [15], [16], [17]. And in 1998, edge-preserving median root prior (MRP) was proposed by Alenius et al. for iterative reconstruction of PET transmission images [18].
However, in the case of short transmission scan when the noise level is relatively significant, edge-preserving nonquadratic priors tend to produce blocky piecewise regions or so-called staircase artifacts. What is more, the application of edge-preserving Bayesian methods are often hindered by the generally needed annealing procedures and complicated parameter estimations for some built-in parameters [10], [15], [16], [17]. Furthermore, the prior energies’ convexities can not be preserved for some edge-preserving nonquadratic priors, which might endanger the whole concavities of their overall objective posterior energy functions and hinder the application of convergent algorithms [10].
The quadratic prior smooth the reconstructed image through an averaging alike operation on pixel densities within a local neighborhood. The edge-preserving nonquadratic priors rely on pixel intensity difference information within a local neighborhood to determine the degrees of regularizing effect on every pixel in image [10], [11], [12], [13], [14], [15], [16], [17]. Both the annoying oversmoothing effect for quadratic priors and staircase effect for nonquadratic priors are the outcomes of their limited power of distinguishing edge information from noise information indiscriminatively. And such limited power can be ascribed to the fact that the simply weighting strategy for pixel density differences within a small fixed local neighborhood can only provide limited prior information. None of above priors addresses the information of global connectivities and continuities in objective image and we term these traditional priors local priors.
Yu et al. devised a boundary-based Bayesian method which incorporates global information of image by level-set methods [19]. But such boundary-based method relies heavily on the level-set operations whose effect in different images is unpredictable and parameter-dependent. Recently, an effective nonlocal MRF quadratic prior model for emission image reconstruction is proposed [20]. This nonlocal quadratic prior can greatly improve the reconstruction by exploiting not only density differences information between individual pixels but also global connectivity and continuity information in the objective image.
In this article, we apply such nonlocal MRF quadratic prior model in transmission image reconstruction. Our aim is to reconstruct high quality transmission images even with relatively short acquisition times. In Section 2, the review of the old local prior model and the theory for the proposed nonlocal prior model are both illustrated. In Sections 3 Joint reconstruction strategy, 4 Experimentations, we give the iterative reconstruction algorithm and perform PET transmission reconstruction using the proposed prior for both simulated data and real scan data. Relevant qualitative and quantitative comparisons show the proposed nonlocal prior’s good properties in lowering noise effect and preserving edges for transmission reconstruction with different scan times. Conclusions and plans for future work are given in Section 5.
Section snippets
Theory of the proposed nonlocal MRF prior model
Because of low counting rates and noise effect, the reconstruction of an unknown objective image f from measurement data g, such as the reconstruction of PET emission and transmission images, is often an ill-posed problem [1], [9], [10]. Bayesian methods, or maximum a posteriori (MAP) ways, have already been accepted as an effective solution to above problem [21], [22]. Based on Bayesian and MRF theory, a generic contextual constraints can be transformed into some kind of prior information to
Statistical model
In PET statistical reconstruction, transmission scan measurement data vectors can be independently modeled as Poisson random variables. , the measurement detected by the detector pair i from a transmission scan, can be well modeled as Poisson random with expectation as a function of the underlying attenuation map for transmission tomography [9], [10], [11], [12], [13], [14]. In transmission tomography, (i.e., the likelihood function) is the probability of obtaining the measurement
Simulated phantom data case
In this experiment, synthetic simulated phantom data with square pixels are used for transmission reconstruction. Fig. 2 shows the simulated 2D PET attenuation map. The attenuation map comprises four Regions which correspond to different substances with assumed attenuation coefficient levels ( cm−1, cm−1, soft tissue cm−1, bone cm−1), and the pixel values in the map are set to be 0 for air, 0.02 for lungs, 0.096 for soft tissue and 0.2 for bone
Conclusions and future work plan
Our motivation lies in the fact that attenuation correction is aimed mainly at providing improved quantification of attenuation correction images in which the organs are significantly different in their attenuation. Therefore, in reconstructing attenuation maps, it is important to preserve the edges that bound anatomical regions. Stemming from MRF theory, the application of the proposed nonlocal prior is theoretically correct and straight forward. Reconstruction using the nonlocal prior can be
Acknowledgment
The authors would like to thank Professor J. A. Fessler of the university of Michigan for software support.
Yang Chen received his BS, MS and PhD degrees in biomedical engineering from the Department of Biomedical Engineering, First Military Medical University, China. He is currently a Lecturer at the Laboratory of Image Science and Technology, Southeast University, Nanjing, China. His research interests include medical imaging reconstruction (especially on PET/CT), and medical image analysis.
References (27)
- et al.
Bayesian statistical reconstruction for low-dose X-ray computed tomography using an adaptive-weighting nonlocal prior
Computerized Medical Imaging and Graphics
(2009) - et al.
Maximum likelihood reconstruction for emission tomography
IEEE Transactions on Medical Imaging
(1982) - et al.
Positron emission tomography
IEEE Signal Processing Magzine
(1997) Transmission scanning in emission tomography
European Journal of Nuclear Medicine
(1998)- et al.
Attenuation correction using count-limited transmission data in positron emission tomography
Journal of Nuclear Medicine
(1993) - et al.
Detailed investigation of transmission and emission data smoothing protocols and their effects on emission images
IEEE Transactions on Nuclear Science
(1996) Hybrid Poisson/polynomial objective functions for tomographic image reconstruction from transmission scans
IEEE Transactions on Image Processing
(1995)- et al.
Adaptive, segmented attenuation correction for whole-body PET imaging
IEEE Transactions on Nuclear Science
(1996) - et al.
A comparison of segmentation and emission subtraction for singles transmission in PET
IEEE Transactions on Nuclear Science
(1998) Maximum-likelihood reconstruction of transmission image in emission computed tomography via the EM algorithm
IEEE Transactions on Medical Imaging
(1990)
Convergence of EM image reconstruction algorithms with Gibbs smoothness
IEEE Transactions on Medical Imaging
Grouped-coordinate ascent algorithms for penalized likelihood transmission image reconstruction
IEEE Transactions on Medical Imaging
Globally convergent algorithms for maximum a posteriori transmission tomography
IEEE Transactions on Image Processing
Cited by (6)
Nonlocal Markovian models for image denoising
2016, Journal of Electronic ImagingPET reconstruction via nonlocal means induced prior
2015, Journal of X-Ray Science and TechnologyDocument image super-resolution using structural similarity and Markov random field
2014, IET Image ProcessingTotal variation deconvolution for terahertz pulsed imaging
2011, Inverse Problems in Science and EngineeringApplication of penalized least-squares algorithm in PET image reconstruction based a nonlocal quadratic prior
2011, Proceedings of SPIE - The International Society for Optical EngineeringEffective image testorations using a novel spatial adaptive prior
2010, Eurasip Journal on Advances in Signal Processing
Yang Chen received his BS, MS and PhD degrees in biomedical engineering from the Department of Biomedical Engineering, First Military Medical University, China. He is currently a Lecturer at the Laboratory of Image Science and Technology, Southeast University, Nanjing, China. His research interests include medical imaging reconstruction (especially on PET/CT), and medical image analysis.
Liwei Hao received the PhD degrees in biomedical engineering from the Department of Biomedical Engineering, First Military Medical University, China. He is currently a Lecturer in the School of Biomedical Engineering and the director of the Key Lab for Medical Image Processing of Guangdong province. His work is on the 3D image reconstruction.
Xianghua Ye is currently a PHD candidate in the Department of Radiation Oncology, Nanfang Hospital Affiliated to Southern Medical University, China.
Wufan Chen received the BS and MS degrees in applied mathematics, computational fluid dynamics from Peking University of Aeronautics and Astronautics (BUAA), China, in 1975 and 1981, respectively. Since September 2004, he has been with Southern Medical University, China, where he holds the rank of Professor in the School of Biomedical Engineering and the director of the Key Lab for Medical Image Processing of Guangdong province. His research focuses on the medical imaging and medical image analysis.
Limin Luo is currently a professor at Laboratory of Image Science and Technology, Southeast University, Nanjing, China. He is also a senior member of IEEE. His research interests include medical imaging instrumentation and biomedical engineering.
Xindao Yin received a medical degree from the University of SuZhou, China, in 1989 and PhD from the University of FuDan, China, in 2002. Since 2003, he has been working in Nanjing First Hospital Affiliated to Nanjing Medical University, where he is an associate professor in radiology and a chief in CT/MRI Department. His main interests are CT or MRI clinical diagnosis and medical image post-processing.
- ☆
This research was supported by National Basic Research Program of China under grant, No. 2010CB732503.