Low-dose computed tomography reconstruction regularized by structural group sparsity joined with gradient prior
Introduction
Nowadays, X-ray computed tomography (CT) has been widely used as one of the best effective imaging diagnosis tools. Since excessive radiation dose delivered to patient might induce high risk of cancer and gene mutation [1,2], the reduction of radiation dose has arisen worldwide concerns. Therefore, the investigation of low-dose computed tomography (LdCT) imaging technique is active in medical field.
To reduce the radiation dose, there are many approaches which can be mainly divided into two categories: one is to change the scanning protocol by lowering the tube voltage and/or tube current [3,4], the other is to down-sample the measured data for CT reconstruction (such as interior CT [5], [6], [7], limited-angle CT [8], [9], [10], [11], [12] and sparse-view CT [13], [14], [15], [16]). However, the radiation dose reduction using the first category will make the measured CT data contaminated by serious noise. The quality of CT image reconstructed by standard filtered back-projection (FBP) method [17] is seriously degraded by the noise-contaminated CT data. As for the second category to reduce the radiation dose, the measured CT data will be truncated for CT reconstruction, which will lead to ill-posed inverse problem. Especially when the measured CT data is incomplete for CT reconstruction, the reconstruction images by FBP method will present terrible stripe artifacts with serious noise and/or limited-angle artifacts. Thus, to maintain clinically acceptable image quality for LdCT is desirable in practical applications. Some existing works have shown that when the radiation dose is significant reduced as in LdCT, iterative reconstruction techniques can provide similar image quality compared with the reconstruction by FBP with normal dose [18], [19], [20], [21], [22].
In this work, we will focus the first category for LdCT. To deal with the problems mentioned above, many scholars have made great efforts in this field. As analytic reconstruction methods (such as FBP) present poor performance for LdCT imaging, classical iterative reconstruction (IR) methods (such as Algebraic Reconstruction Technique [23], ART and Simultaneous Algebraic Reconstruction Technique [24], SART) combined with prior regularization constraint proved to be effective. Total variation (TV) [25] known as one of the famous prior regularization constraints has been widely used for CT reconstruction with under-sampled projection data [9,26]. Since the limitation of TV regularization lies in the assumption that the reconstruction image is piecewise constant [25], the reconstructed image by TV may loss some tiny details and texture features. Therefore, to improve the image quality, its variants for LdCT have been proposed [27], [28], [29], [30], [31], [32], [33], [34]. Tian etal [27] designed an edge-preserving TV norm as regularization term which is prefer to smoothing the non-edge part of the object and giving low weight to penalize the pixels at the edges. Liu etal [28] designed a local image-intensity gradient dependent weight for TV minimization which considers the anisotropic edge property among neighboring image voxels. Yang etal [29] generalized the piecewise constant assumption of TV minimization to piecewise polynomial, and then introduced high-order TV minimization for interior problem. Ritschl etal [30] proposed an improved method to optimize the parameter adaption within the framework of [26]. This improvement makes their algorithm converge to the lowest possible value of object function and meanwhile hold a low value of the sparsity constraint. As the good performance of total generalized variation (TGV) [31] regularization which utilizes higher order derivatives of the image to be reconstructed, Niu etal [32] presented a penalized weighted least-squares (PWLS) scheme based on TGV and showed the superiority of their method in accuracy and resolution properties. Analogously, to penalize the higher-order derivatives of the image to be reconstructed, Zeng etal [33,34] utilized structure tensor total variation (STV) regularization to eliminate the patchy artifacts usually appeared in TV regularization for multienergy CT and Cerebral perfusion CT. Niu etal [35] considered the low-rank and sparse matrix characteristic of a series of enhanced sequential perfusion X-ray CT images, and then proposed a low-dose PCT image restoration model to reduce the radiation exposure. Since the measured projection data is degraded with serious noise, some sinogram filtering methods are feasible to estimate noise-free sinogram for CT reconstruction [36], [37], [38]. Zhang etal [38] proposed a low-dose CT scheme which includes removing isolated data in sinogram domain and exploited segmentation-based adaptive filtering technique. Lee etal [39] demonstrated that sinogram-affirmed iterative reconstruction enables ultra-low-dose CT with very low radiation doses. Wang etal [40] analyzed the noise property of CT sinogram data in Radon space with different mAs levels which is helpful to improve CT image quality for screening applications. Some other scholars considered incorporating non-local means image filtering [41,42]. Ma etal [41] maked full use of the redundancy information from previous normal-dose scan by which the non-local weights are calculated. Instead of replacing previous normal-dose scan, Bian etal [42] used the FBP image reconstructed from the sinogram restored by PWLS algorithm in the Karhunen-Loéve domain. The presented algorithm makes full use of the advantages of the restored sinogram and non-local means image filtering algorithm and shows superiority in terms of noise reduction and image resolution. Considering both the X-ray photon statistics and the electronic noise background, Xie etal [43] formulated a robust sinogram preprocessing based on noise-generating mechanism. In order to evaluate the effect on the radiation exposure and image quality, Cianci etal [44] adopted a scheme which combined automatic tube current modulation and sinogram-affirmed iterative reconstruction technique in ultra-low dose CT colonography. Compared with low-dose CT colonography, their works showed that the radiation dose in ultra-low dose CT colonography can be reduced up to 63% and there is no relevant image quality deterioration. Considering the similarity of spectral CT images, a reconstruction method combined image gradient l0-norm constraint and tensor dictionary learning (TDL) for spectral CT was proposed [45], whose method emphasizes the spatial sparsity which can overcome the drawback of TDL on edge preserving. Considering the good property of relative total variation (RTV), Gong etal [46] proposed RTV-based method for CT reconstruction, the structural information of the target very well, however, details are easy to be lost. With the aid of previous high-dose reconstructions of the same object, templates technology can also play an important role in CT reconstruction, which can compensate for the low-dose artefacts [47].
In recent years, machine learning technology has been extensively studied. As an important branch of machine learning, deep learning (DL) [48] technique especially convolutional neural network (CNN) based method demonstrates superior performance in the field of image processing such as image classification, image detection, image segmentation and image restoration. Due to the good performance of DL-based method in these areas, some scholars have extended it to medical imaging domain such as LdCT for high precision CT reconstruction [49], [50], [51], [52], [53], [54], [55]. The success of these DL-based method for LdCT benefits from a large number of data training samples which is significantly important for deep learning. It is well known that DL-based methods usually require high computational power and storage capacity which depend on computer hardware devices, such as Graphics Processing Unit (GPU) etc. However, this requires the upgrading of existing computer hardware which would require additional funding. Due to the inconsistent storage standards of most CT manufacturers’ original CT data and the confidentiality of patients’ CT data, it is difficult to meet the training sample size required for DL-based CT reconstruction methods, which limits the promotion of DL technology. In addition, DL-based methods are sensitive to hyperparameters which need a long time (usually measured in hours or days) to be chosen.
Since iterative reconstruction method is more flexible and advantageous than analytic method, in this paper, we will study iterative reconstruction algorithm for LdCT without the help of huge CT image database. As discussed above, although the gradient-based sparsity in image domain denoted as TV is one of the popular prior information for CT reconstruction, there is some shortcomings presented in TV-based results. In this study, inspired with the structural group sparse representation (SGSR) proposed for natural image restoration [56], we present a hybrid regularization constraint for LdCT reconstruction problem. The presented CT reconstruction method can overcome the shortcomings of TV-based reconstruction method since both the structural group sparsity and gradient sparsity (SGSaGS) are simultaneously enforced as the regularization constraint in image domain. To effectively solve the proposed unconstrained optimization problem, we utilize the alternating direction method of multipliers (ADMM) and then transformed the original optimization problem into a series of sub-problems. Both the simulated phantom study and real CT data study performed to qualitatively and quantitatively evaluate the proposed LdCT reconstruction. The results show that the proposed algorithm has the potential to spectral computed tomography.
The remainder of this paper is organized as follows. In Section 2, the CT imaging model, gradient sparsity-based CT reconstruction model and brief review of structural group sparse (SGS) representation modeling are respectively presented, and then the proposed CT reconstruction model is described in detail. In Section 3, our experimental results of LdCT image reconstruction are provided. Finally, discussion and conclusion are given in Section 4.
Section snippets
CT imaging model
Low-dose CT whose purpose is to reconstruct high quality CT image from noisy degraded data, is a typical ill-posed linear inverse problem. In the real situation, CT imaging model can be approximately formulated as:where denotes the CT system matrix or projection operator, denotes the measured CT projection data degraded by the noise level and denotes the linear attenuation coefficients of the object to be reconstructed; is the number of detector bins and is the
Results
In this section, we present the performance of the proposed algorithm for LdCT from simulated projection data and real data. All the experiments were implemented on a desktop in MATLAB 2015b combined with Microsoft VC++ 2010. The following experiment takes fan-beam CT scanning as an example. As for simulated projection data experiments, different levels of Poisson noise were considered. As for real projection data experiments, two levels of Poisson noise were added to the real CT projection
Discussion and conclusion
In this work, we investigate regularized reconstruction algorithm for low-dose CT reconstruction without the aid of measured high-dose CT data. Compared with other reconstruction algorithms, such as SART, ADMM-TV and RTV, the proposed method also can well protect the details and edge structure for CT imaging with different low-dose levels. The results by RTV based method in reference [56] show that he structural information of the target is well preserved while the tiny details and textures may
CRediT authorship contribution statement
Wei Yu: Writing - original draft, Writing - review & editing, Methodology. Wei Peng: Methodology, Formal analysis. Hai Yin: Software, Validation. Chengxiang Wang: Writing - review & editing, Supervision. Kaihu Yu: Conceptualization.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgments
The authors declare that there are no conflicts of interest related to this article. This work was partially funded by the National Natural Science Foundation of China (No. 61701174 and 61801086); Natural Science Foundation of Hubei Province (No. 2017CFB168); Natural Science Foundation of Chongqing (No. cstc2019jcyj-msxmX0345); Scientific Research Founds of Chongqing Normal University (No. 19XLB005). The authors also thank Guangzhou Huaduan Technology co., LTD for providing real data for CT
References (61)
- et al.
A variational proximal alternating linearized minimization in a given metric for limited-angle CT image reconstruction
Appl. Math. Modell.
(2019) - et al.
System matrix analysis for sparse-view iterative image reconstruction in X-ray CT
J. X-ray Sci. Technol.
(2015) - et al.
Radiation dose of non-enhanced chest CT can be reduced 40% by using iterative reconstruction in image space
Clin. Radiol.
(2011) - et al.
Nonlinear total variation based noise removal algorithms
Phys. D: Nonlin. Phenom.
(1992) - et al.
Penalized weighted least-squares approach for multienergy computed tomography image reconstruction via structure tensor total variation regularization
Comput. Med. Image. Graph.
(2016) - et al.
Low-dose cerebral perfusion computed tomography image restoration via low-rank and total variation regularizations
Neurocomputing
(2016) - et al.
SR-NLM: a sinogram restoration induced non-local means image filtering for low-dose computed tomography
Comput. Med. Image. Graph.
(2013) - et al.
Low-dose spectral CT reconstruction using image gradient l(0)-norm and tensor dictionary
Appl. Math. Modell.
(2018) - et al.
Adaptive iterative reconstruction based on relative total variation for low-intensity computed tomography
Signal Process.
(2019) - et al.
Low radiation tomographic reconstruction with and without template information
Signal Process.
(2020)
Structure extraction from texture via relative total variation
ACM Trans. Graph.
Radiation dose associated with common computed tomography examinations and the associated lifetime attributable risk of cancer
Arch. Internal Med.
Cancer risks from diagnostic radiology
Br. J. Radiol.
Low-tube voltage 100 kVp MDCT in screening of cocaine body packing: image quality and radiation dose compared to 120 kVp MDCT
Abdom. Image.
Feasibility study of radiation dose reduction in adult female pelvic CT scan with low tube-voltage and adaptive statistical iterative reconstruction
Korea. J. Radiol.
Compressed sensing based interior tomography
Phys. Med. Biol.
High-order total variation minimization for interior tomography
Inverse Prob.
The meaning of interior tomography
Phys. Med. Biol.
Wavelet-based reconstruction for limited-angle X-ray tomography
IEEE Trans. Med. Image.
Accurate image reconstruction from few-views and limited-angle data in divergent-beam CT
J. X-ray Sci. Technol.
A limited-angle CT reconstruction method based on anisotropic TV minimization
Phys. Med. Biol.
Edge-preserving reconstruction from sparse projections of limited-angle computed tomography using ℓ0-regularized gradient prior
Rev. Sci. Instrum.
Sparse tomography
SIAM J. Sci. Comput.
Low-dose X-ray computed tomography image reconstruction with a combined low-mAs and sparse-view protocol
Opt. Exp.
Sparse-view image reconstruction via total absolute curvature combining total variation for X-ray computed tomography
J. X-ray Sci. Technol.
The Mathematics of Computerized Tomography
CT radiation dose and iterative reconstruction techniques
Am. J. Roentgenol.
CT pulmonary angiography: dose reduction via a next generation iterative reconstruction algorithm
Acta Radiol.
Chest computed tomography using iterative reconstruction vs filtered back projection (Part 2): image quality of low-dose CT examinations in 80 patients
Eur. Radiol.
Adaptive statistical iterative reconstruction technique for radiation dose reduction in chest CT: a pilot study
Radiology
Cited by (8)
An anisotropic alternating regularization-based reconstruction algorithm for cone beam computed laminography
2023, NDT and E InternationalNonconvex L<inf>1/2</inf>- regularized nonlocal self-similarity denoiser for compressive sensing based CT reconstruction
2023, Journal of the Franklin InstituteSparse-View Computed Tomography Reconstruction Based on a Novel Improved Prior Image Constrained Compressed Sensing Algorithm
2023, Applied Sciences (Switzerland)Deep Dual-Domain United Guiding Learning With Global-Local Transformer-Convolution U-Net for LDCT Reconstruction
2023, IEEE Transactions on Instrumentation and MeasurementProjection domain processing for low-dose CT reconstruction based on subspace identification
2023, Journal of X-Ray Science and Technology