Scatter correction based on adaptive photon path-based Monte Carlo simulation method in Multi-GPU platform
Introduction
Since the first CT was developed by G. N. Hounsfield in 1972, CT has been rapidly applied in lots of fields especially in medical clinic with its great advantages of high resolution and sensitivity. CT plays an irreplaceable role in medical diagnosis [1]. In the late 1990s, Cone-beam CT (CBCT) appeared, making medical CT imaging technology of epoch-making significance. However, since CBCT cannot use the collimator technology like spiral CT [2], there are serious scatter artifacts [3], which has seriously restricted the in-depth application and development of CBCT in clinical practice.
Many researches have been carried out on the correction of scatter artifacts. Boone J.M concluded it as two types of scatter correction technology based on hardware and software [4]. Methods based on hardware mainly contain anti-scatter grids, air gap [5], collimator, etc. In addition, primary modulation methods [6,7] are proposed in the Fourier space to extract the low-frequency scatter signal. Furthermore, patient planning CT has also been adapted as the prior information for scatter estimation [8]. Nonetheless, since the hardware-based approach achieves its goal in the way of X-ray loss, it also brings some problems, such as reduction for signal-to-noise ratio (SNR) of image, increase of patient radiation dose and so on [4,6].
Two main representative methods based on software are the scatter kernel deconvolution [9,10] and MC simulation. The scatter kernel deconvolution is not accurate enough because of the uniqueness and linearity of kernel convolution. Calculation method based on MC simulation is known as the golden standard for particle physical transport in radiation therapy [11]. In image-guided radiation therapy (IGRT), the conventional MC simulation takes about more than 1 minute.
It is worth mentioning that the graphics processing unit (GPU) [12,13] can also significantly boost the efficiency. Jarry G et al. [3] proposed a comprehensive framework of artifact correction for CBCT, which was proved to increase the efficiency by several orders of magnitude [14,15]. Some techniques such as VRTs, reductions of particles and projections, projection de-noising and so on can accelerate those GPU-based MC scatter correction methods. However, those conventional MC simulation inevitably uses the scatter-contaminated CBCT images to improve the calculation accuracy through the iteration [16,17], which makes the HU value inaccurate and substantially extends the simulation time.
For solving the problem, a planning CT image, typically available for cancer treatment, is incorporated in our MC simulations. In the previous method, a combination of several techniques including scatter de-noising and image down-sampling are utilized to substantially accelerate the calculation of scatter estimation [18]. In our previous paper, we adopt accelerated strategies including reducing photon history number, pixels sampling, projections sampling and phantom image down-sampling to achieve adaptive fast CBCT image reconstruction. With all the strategies used, the computation time within 15 seconds is achieved for scatter estimation of one CBCT case in IGRT. We found that conventional MC methods even in GPU platform cannot meet the time requirement in the clinical environment. In reality, seconds or shorter are desired to expand the clinical application [19,20]. To solve the problem, we utilize an original GPU-based Metropolis MC (gMMC) scheme for scatter simulation. It significantly accelerate the convergence by adapting a Metropolis algorithm with a path-by-path sampling method to artificially control each path of photon deposited in the detector [21]. In this paper, we demonstrate that the proposed scatter correction workflow based on gMMC and sparse simulation techniques, which calculate the scatter simulation of half-fan and full-fan CBCT cases on the 4-GPU platform. We can achieve computation time within 2.5s for scatter simulation and 15s for the whole workflow of scatter correction and image reconstruction for one CBCT case.
Section snippets
Method and materials
The workflow of the scheme is shown in Fig. 1. The whole workflow including 7 steps and all acceleration strategies are accomplished on a 4-GPU computer workstation to achieve high computation. First, the original projection is adopted to be reconstructed by the Feldkamp-Davis-Kress (FDK) algorithm [22]. The reconstructed image as the initial CBCT image contains scatter artifacts. It is rigidly registered with the planning CT image in the form of convolution. Second, the image after the rigid
Parameter Analysis in the Sparse Simulation
As described in Section 2.2.1, we preliminarily concludeas (Np, Ns,Nt,Nd) (107, 4, 4, 12). Then, we keep three parameters unchanged in the following trials, adjusting the other parameter, and tested whether it meets the required conditions.
Discussions
In the proposed method, gMMC, an original Markov Chain Monte Carlo-based algorithm is implemented for scatter estimation. In contrast with classical MC simulation, the characteristic rationale of gMMC indicated that the proposed scheme can be sped up by sampling the pixels with the appropriate interval. The whole workflow including 7 primary steps and other acceleration strategies is ran on a 4-GPU platform to achieve a high computational efficiency. The phantom and clinical cases show the
Conclusions
In this study, by the advantage of the sampling scheme in gMMC, sparsely selecting the photon deposition pixels in the projection can achieve MC computation of scatter estimation within 2.5s and the total time of CBCT reconstruction with scatter correction less than 15s in a 4-GPU workstation. The time cost of algorithm program is comparable to the scatter kernel deconvolution method. The proposed algorithm is fast enough for CBCT imaging task in the IGRT and dental clinic.
Declaration of Competing Interest
None.
Acknowledgments
This work was supported by the National key R&D Program of China (2016YFA0202003), National Natural Science Foundation of China under Grant (81601493 and 61971463), Ministry of Science and Technology Planning Project of Guangdong (2015B020233005; 2015B020233002); Guangdong Natural Science Foundation of China (2016A030310388 and 2017A030313692); Pearl River Nova Program of Guangzhou (201906010013; 201906010014)
References (32)
- et al.
High-performance cone beam reconstruction using CUDA compatible GPUs
Parallel Computing
(2010) - et al.
“Statistical image reconstruction for low-dose CT using nonlocal means-based regularization
Computerized Medical Imaging and Graphics
(2014) - et al.
Perfusion CT of the brain: diagnostic approach for early detection of ischemic stroke
Radiology
(1998) Computed tomography: principles, design, artifacts, and recent advances
- et al.
Characterization of scattered radiation in kV CBCT images using Monte Carlo simulations
Medical Physics
(2006) Scatter correction algorithm for digitally acquired radiographs: Theory and results
Medical Physics
(1986)Grids or air gaps for scatter reduction in digital radiography: a model calculation
Medical Physics
(1992)- et al.
Scatter Correction Method for X-Ray CT Using Primary Modulation: Theory and Preliminary Results
IEEE Transactions on Medical Imaging
(2006) - et al.
Modulator design for x-ray scatter correction using primary modulation: Material selection
Medical Physics
(2010) - et al.
Quantitative cone-beam CT imaging in radiation therapy using planning CT as a prior: First patient studies
Medical Physics
(2012)
Efficient scatter correction using asymmetric kernels
Physics of Medical Imaging
Improved scatter correction using adaptive scatter kernel superposition
Physics in medicine and biology
Fifty years of Monte Carlo simulations for medical physics
Physics in medicine and biology
Clinical applications of cone-beam computed tomography in dental practice
Journal of the Canadian Dental Association
Computed Tomography: Principles, Design, Artifacts, and Recent Advances
SPIE—the International Society for Optical Engineering
Accelerating Monte Carlo simulations of photon transport in a voxelized geometry using a massively parallel graphics processing unit
Medical Physics
Cited by (0)
- 1
contributed equally