連祥媛,孔慧華,潘晉孝*,高文波,王 攀
多通道聯(lián)合的廣義總變分能譜CT重建
連祥媛1,2,孔慧華1,2,潘晉孝1,2*,高文波1,2,王 攀3
1中北大學理學院,山西 太原 030051;2信息探測與處理山西省重點實驗室,山西 太原 030051;3湖南云箭集團有限公司,湖南 辰溪 419503
基于光子計數(shù)探測器的能譜CT在材料分解、組織表征、病變檢測等應用中具有巨大的潛力。在重建過程中,通道數(shù)的增加會造成單通道中光子數(shù)減少,從而導致重建圖像質(zhì)量下降,難以滿足實際需求。本文從能譜CT重建的角度出發(fā),將廣義總變分向矢量延伸,利用奇異值的稀疏性,促進圖像梯度的線性依賴,提出一種基于核范數(shù)的多通道聯(lián)合廣義總變分的能譜CT重建算法。在圖像重建過程中,多層共享結構信息,同時保留獨特的差異。實驗結果表明,本文提出的算法在抑制噪聲的同時,能夠更有效地恢復圖像細節(jié)及邊緣信息。
CT重建;能譜CT;廣義總變分;核范數(shù);多通道聯(lián)合
在基于光子計數(shù)探測器(photon-counting detector,PCD)的能譜CT[1]中,PCD技術消除了電子噪聲,提供了比傳統(tǒng)CT探測器更高的信噪比[2]。然而,PCD技術仍然存在兩個主要的問題,首先,單個能量通道只包含總光子的一小部分;其次,大部分PCD只能承受有限的計數(shù)率,所以從PCD獲得的多通道投影通常包含非常強的泊松噪聲。因此,開發(fā)能譜CT重建算法對于改善臨床應用具有重要意義[3]。
壓縮感知理論的出現(xiàn),使正則化約束項的迭代類重建算法發(fā)揮出巨大的潛力。能譜CT重建中,正則化約束項以先驗圖像引導和稀疏性條件為主。先驗圖像引導以Yu采用全光譜圖像作為先驗圖像的壓縮感知(prior image constrained compressed sensing,PICCS)算法[4]為代表,稀疏性條件[5-8]包括全變分、緊框架、小波和字典學習等已被應用于CT重建,并取得了不同程度的成功。受空間域信息相關性的啟發(fā),Zhang等[9]將總變分與能譜均值相結合,提高重建圖像的質(zhì)量。Li等[10]提出一種能譜非局部均值的方法,利用圖像相關性來抑制噪聲和條紋偽影。Hu等[11]利用張量字典塊的稀疏性表示來提高重建質(zhì)量。陳佩君等[12]提出一種總變分與傳統(tǒng)張量字典學習結合的重建算法,可進一步恢復圖像的微小結構,有效抑制噪聲。為了更好地利用能譜CT在能量軸方向的信息相關性,Rigie等[13]將總變分向矢量延伸,提出了總核變分(total nuclear variation,TNV)正則化方法,可以更好地保存圖像特征。Niu等[14]提出了一種alpha發(fā)散約束的廣義總變分(total generalized variation,TGV)方法用于稀疏視角X射線CT圖像重建,有效地消除總變分正則化中經(jīng)常出現(xiàn)的階梯狀和斑片狀偽影。
為了更好地使用空間的結構信息,本文將單通道的TGV向矢量延伸,提出一種基于核函數(shù)的多通道聯(lián)合TGV的能譜重建算法,簡稱Mutli-NTGV。利用核范數(shù)作為約束,增強通道間耦合,恢復圖像的結構特征。
當兩幅圖像擁有相同的曲線時,兩幅圖像具有相同方向的梯度,反之亦然[16]。如果各通道圖像的所有梯度向量是平行的或反平行的,那么雅可比矩陣的秩將是1,因此只有一個非零奇異值。在此基礎上,若各通道間圖像梯度平行,會使得核范數(shù)最小,將TGV推向矢量場形式如下:
類似地:
其中:
矢量場的離散梯度算子為
對稱化梯度算子為
若:
則:
在能譜CT的不同能量通道下,圖像的梯度信息是相似的。為了更好地克服基于導數(shù)的正則化方法的缺陷,有效地利用通道間的結構化信息。提出一種基于核范數(shù)的多通道聯(lián)合TGV能譜CT重建算法,其目標函數(shù)可以表述為以下凸極小化問題:
式(6)中含有兩個變量,可以采用交替迭代的方法優(yōu)化,將上述式子分裂成兩個子問題:
重新整理正則項的形式,將式(9)的最小化問題表述為
采用一階原始—對偶算法解決上述鞍點問題。
具體的迭代步驟為
在此過程中,青少年價值觀與道德判斷將會得到進一步提升,出現(xiàn)攻擊行為的可能性也會大大降低。與此同時,群體觀與自我人際價值感之間的關系主要呈現(xiàn)負相關關系。主要是指越是認為自我觀重要的個體,個體存在的人際價值感愈加淡薄,相反則反之??梢哉f,學校方面必須加強對青少年價值觀與道德判斷方面的教育能力,盡量規(guī)避青少年攻擊行為的出現(xiàn)。為了進一步提升青少年價值觀與道德判斷能力,教師可以適當提升關于這方面的教育能力,確保青少年規(guī)范自身行為。
其中:離散梯度算子的逆算子為
對稱化算子的逆算子為
能譜的Mutil-NTGV算法步驟:
為驗證本文算法的有效性,采用濾波反投影重建算法(FBP)、聯(lián)合代數(shù)重建算法(SART)、總變分正則項的迭代重建算法(TV)、基于張量的字典學習算法(TDL)、廣義總變分正則化的迭代重建算法(TGV)為比較算法。所用算法都是用MATLAB和C++的混合模式實現(xiàn)的,接口在MATLAB中實現(xiàn),所有的大規(guī)模計算部分在C++中實現(xiàn),并通過MEX函數(shù)進行編譯。本文采用峰值信噪比(PSNR)、歸一化均方根誤差(NRMSE)與結構相似度指數(shù)(SSIM)定量評價各方法性能,驗證算法的有效性。
圖1為FBP、SART、TV、TDL、TGV與Multi-NTGV算法迭代30次的重建效果圖。
從圖1可以看出,F(xiàn)BP、SART重建的結果中噪聲很強,TDL、TGV與Mutli-NTGV都可以不同程度地減少各個能譜通道下噪聲的影響。在重建效果的對比中,可以明顯地看出TV在平滑噪聲的過程中,容易產(chǎn)生階梯狀偽影,導致重建效果不佳。TDL平滑效果較好,但無法區(qū)別噪聲與細節(jié),局部去噪效果不佳。TGV很好地克服了TV中所產(chǎn)生的階梯偽影的狀況,保存更多的細節(jié),但在邊緣部分會產(chǎn)生一點弱化的效果。本文提出的方法利用了通道間的結構化信息,不僅有效地抑制了噪聲,而且對于微小的結構也保持得很好,邊緣更加清晰。
五種算法重建過程中的NRMSE、PSNR與SSIM指標如圖2所示,F(xiàn)BP重建效果的評價指標如表1所示。由圖2可以看出,在前三個通道中,本文所提出的算法明顯優(yōu)于其他算法,在第四通道中,雖然Mutli-NTGV算法與TDL、TGV的數(shù)量性評價結果相近,但是從重建圖效果可以看出Mutli-NTGV方法優(yōu)于其他方法。實驗結果表明,本文所提的方法在各個通道中,會保持較大的優(yōu)勢,極大地提升了重建圖像的質(zhì)量。
為了進一步驗證所提出方法的有效性,采用了MARS(Medipix All Resolution System)微型CT上采集的來自真實臨床前小鼠的投影。電壓設置為120 kV,電流為175 mA。從照射源到系統(tǒng)原點的距離為158 mm,到探測器的距離為255 mm。在整個掃描范圍內(nèi)均勻收集了13個能量通道的371個投影視圖。探測器一行1024個元素,單位長度為55mm。
圖3為各算法對臨床前小鼠的重建效果圖,從上到下依次展示的是第1、3、5和7通道的效果圖。實際小鼠切片實驗中,各算法重建結果與小鼠胸腔仿真實驗類似,F(xiàn)BP與SART重建效果中含有大量的噪聲,其余方法都在一定程度上達到了去噪的效果,從實驗結果可以明顯看出,在各個重建算法中,Mutli-NTGV在臨床小鼠的重建效果較好。
圖1 小鼠模型的重建結果。從左到右的方法依次為FBP、SART、TV、TDL、TGV與Mutli-NTGV,從上到下依次為1到4通道
圖2 小鼠模型重建效果的數(shù)量性評價指標。 從左到右行依次為NRMSE、PSNR、SSIM,從(a)到(d)行依次為1到4通道
表1 FBP重建的數(shù)量性評價指標
圖3 臨床小鼠模型的重建結果。從左到右的方法依次為FBP、SART、TV、TDL、TGV與Mutli-NTGV,從上到下依次為1, 3, 5, 7通道
本文針對能譜CT,提出了一種基于核范數(shù)的多通道聯(lián)合廣義總變分能譜CT重建算法,將TGV推向矢量化,采用逐像素的更新方式,使用多通道聯(lián)合的二階廣義總變分作為專用的正則化函數(shù)在圖像重建步驟中耦合多個通道,耦合分別在一階和二階導數(shù)的水平上用核范數(shù)和F范數(shù)約束來實現(xiàn),促進了多個圖像通道的梯度的線性依賴,從而使邊緣對齊。該方法在噪聲條件下,可以保持更多的細節(jié)與更清晰的邊緣。仿真數(shù)據(jù)與實際數(shù)據(jù)的運行結果表明,本文提出的算法具有較好的魯棒性。但是,本文算法中參數(shù)較多,采用經(jīng)驗選取,比較耗費時間,后續(xù)研究中可以研究采用更靈活的參數(shù)選取模型。
[1] Niu S Z, Bian Z Y, Zeng D,. Total image constrained diffusion tensor for spectral computed tomography reconstruction[J]., 2019, 68: 487–508.
[2] Taguchi K, Iwanczyk J S. Vision 20/20: Single photon counting X-ray detectors in medical imaging[J]., 2013, 40(10): 100901.
[3] Dong X, Niu T Y, Zhu L. Combined iterative reconstruction and image-domain decomposition for dual energy CT using total-variation regularization[J]., 2014, 41(5): 051909.
[4] Yu Z C, Leng S, Li Z B,. Spectral prior image constrained compressed sensing (spectral PICCS) for photon-counting computed tomography[J]., 2016, 61(18): 6707–6732.
[5] Zhang W K, Zhang H M, Wang L Y,. Limited angle CT reconstruction by simultaneous spatial and Radon domain regularization based on TV and data-driven tight frame[J]., 2018, 880: 107–117.
[6] Luo X Q, Yu W, Wang C X. An image reconstruction method based on total variation and wavelet tight frame for limited-angle CT[J]., 2018, 6: 1461–1470.
[7] Us D, Ruotsalainen U, Pursiainen S. Combining dual-tree complex wavelets and multiresolution in iterative CT reconstruction with application to metal artifact reduction[J]., 2019, 18: 116.
[8] Miao J Y, Cao H L, Jin X B,. Joint sparse regularization for dictionary learning[J]., 2019, 11(5): 697–710.
[9] Zhang Y, Xi Y, Yang Q S,. Spectral CT reconstruction with image sparsity and spectral mean[J]., 2016, 2(4): 510–523.
[10] Li B, Shen C Y, Chi Y J,. Multienergy cone-beam computed tomography reconstruction with a spatial spectral nonlocal means algorithm[J]., 2018, 11(2): 1205–1229.
[11] Hu D L, Wu W W, Xu M R,. SISTER: spectral-image similarity-based tensor with enhanced-sparsity reconstruction for sparse-view multi-energy CT[J]., 2020, 6: 477–490.
[12] Chen P J, Feng P, Wu W W,. Material discrimination by multi-spectral CT based on image total variation and tensor dictionary[J]., 2018, 38(11): 1111002.
陳佩君, 馮鵬, 伍偉文, 等. 基于圖像總變分和張量字典的多能譜CT材料識別研究[J]. 光學學報, 2018, 38(11): 1111002.
[13] Rigie D S, Patrick J L R. Joint reconstruction of multi-channel, spectral CT data via constrained total nuclear variation minimization[J]., 2015, 60(5): 1741–1762.
[14] Niu S Z, Huang J, Bian Z Y,. Iterative reconstruction for sparse-view X-ray CT using alpha-divergence constrained total generalized variation minimization[J]., 2017, 25(4): 673–688.
[15] Kristian B, Karl K, Thomas P. Total Generalized Variation[J]., 2010, 3(3): 492–526.
[16] Ehrhardt M J, Arridge S R. Vector-valued image processing by parallel level sets[J]., 2014, 23(1): 9–18.
[17] Sidky E Y, J?rgensen J H, Pan X C. Convex optimization problem prototyping for image reconstruction in computed tomography with the Chambolle–Pock algorithm[J]., 2012, 57(10): 3065–3091.
Joint multi-channel total generalized variational algorithm for spectral CT reconstruction
Lian Xiangyuan1,2, Kong Huihua1,2,Pan Jinxiao1,2*, Gao Wenbo1,2, Wang Pan3
1School of Science, North University of China, Taiyuan, Shanxi 030051, China;2Shanxi Key Laboratory of Signal Capturing & Processing, North University of China, Taiyuan, Shanxi 030051, China;3Hunan Vsngusrd Group. Co. Ltd, Chenxi, Hunan 419503, China
The reconstruction results of the mouse model by Mutli-NTGV
Overview:Spectral computed tomography (CT) based on photon-counting detectors, has great potential in material decomposition, tissue characterization, lesion detection, and other applications. During the reconstruction, the increase of the number of channels will reduce the photon number in a single channel, resulting in the decline of the quality of the reconstructed image, which is difficult to meet the actual needs. To improve the quality of image reconstruction, this paper proposes a joint multi-channel total generalized variational based on the unclear norm for spectral CT reconstruction. Firstly, in the reconstruction for spectral CT, the image structure of each channel is highly similar, and the reconstruction of a single channel will ignore the structural information of each channel. Second, gradient information contains a lot of structured information and features of the image. When two images have the same curve, the two images have the same direction gradient and the converse is also true. In order to better utilize the image’s structural information between channels, the new regularization function is applied to spectral CT reconstruction. The research shows that if the edges of the two images are aligned, the two images have the same gradient. The image gradients between channels are parallel, which will minimize the nuclear norm. The algorithm will extend total generalized variation to the vector, with the aim of overcoming defects of existing derivative-based regularization. The paper proposed a joint multi-channel total generalized variational for spectral CT reconstruction, employing a vectorial second-order total generalized variation function as joint regularization. The method adopts pixel-by-pixel updating in the image reconstruction, and the multi-channel image coupling is realized by kernel norm and F-norm constraints at the level of first and second derivatives. The nuclear norm and frobenius norm coupling promote joint sparsity of the edge sets and dependence of the gradients. Joint multi-channel total generalized variational is used to promote the linear dependence of the multi-channel image’s gradient so that the image edges of each channel are aligned. The structural information of the multi-channel image is shared during the image reconstruction process while unique differences are preserved. The experiment was done on a numerical mouse thorax phantom and clinical mouse data. The quantitative results of peak signal to noise ratio (PSNR), normalized root mean square error (NRMSE) and structure similarity index (SSIM) show that the proposed algorithm greatly improves the image quality. Experimental results show that the proposed algorithm can effectively recover image details and marginal information while suppressing noise.
Lian X Y, Kong H H,Pan J X,Joint multi-channel total generalized variational algorithm for spectral CT reconstruction[J]., 2021, 48(9): 210211; DOI:10.12086/oee.2021.210211
Joint multi-channel total generalized variational algorithm for spectral CT reconstruction
Lian Xiangyuan1,2, Kong Huihua1,2,Pan Jinxiao1,2*, Gao Wenbo1,2, Wang Pan3
1School of Science, North University of China, Taiyuan, Shanxi 030051, China;2Shanxi Key Laboratory of Signal Capturing & Processing, North University of China, Taiyuan, Shanxi 030051, China;3Hunan Vsngusrd Group. Co. Ltd, Chenxi, Hunan 419503, China
Spectral computed tomography (CT) based on photon-counting detectors, has great potential in material decomposition, tissue characterization, lesion detection, and other applications. During the reconstruction, the increase of the number of channels will reduce the photon number in a single channel, resulting in the decline of the quality of the reconstructed image, which is difficult to meet the actual needs. To improve the quality of image reconstruction, joint multi-channel total generalized variational based on the unclear norm for spectral CT reconstruction was proposed in this paper. The algorithm will extend total generalized variation to the vector, and the sparsity of singular values is used to promote the linear dependence of the image gradient. The structural information of the multi-channel image is shared during the image reconstruction process while unique differences are preserved. Experimental results show that the proposed algorithm can effectively recover image details and marginal information while suppressing noise.
CT reconstruction; spectral CT; total generalized variation; nuclear norm; joint multi-channel
連祥媛,孔慧華,潘晉孝,等. 多通道聯(lián)合的廣義總變分能譜CT重建[J]. 光電工程,2021,48(9): 210211
Lian X Y, Kong H H,Pan J X,Joint multi-channel total generalized variational algorithm for spectral CT reconstruction[J]., 2021, 48(9): 210211
10.12086/oee.2021.210211
TP391
A
2021-06-23;
2021-08-26
國家自然科學基金資助項目(61801437,61871351,61971381)
連祥媛(1994-),女,碩士研究生,主要從事圖像重建與圖像處理方面的研究。E-mail:1393550566@qq.com
潘晉孝(1966-),男,博士,教授,主要從事信息處理與圖像重建方面的研究。E-mail:panjx@nuc.edu.cn
National Natural Science Foundation of China (61801437, 61871351, 61971381)
* E-mail: panjx@nuc.edu.cn