吉林大学学报(工学版) ›› 2020, Vol. 50 ›› Issue (5): 1755-1764.doi: 10.13229/j.cnki.jdxbgxb20190554

• 计算机科学与技术 • 上一篇    

基于生成对抗网络的低剂量能谱层析成像去噪算法

史再峰1,2(),李金卓1,曹清洁3,李慧龙1,胡起星1   

  1. 1.天津大学 微电子学院,天津 300072
    2.天津市成像与感知微电子技术重点实验室,天津 300072
    3.天津师范大学 数学科学学院,天津 0087
  • 收稿日期:2019-06-03 出版日期:2020-09-01 发布日期:2020-09-16
  • 作者简介:史再峰(1977-),男,副教授,博士.研究方向:影像感知与智能计算.E-mail:shizaifeng@tju.edu.cn
  • 基金资助:
    国家自然科学基金项目(61674115);天津市自然科学基金项目(17JCYBJC15900)

Low⁃dose spectral computer⁃tomography imaging denoising method via a generative adversarial network

Zai-feng SHI1,2(),Jin-zhuo LI1,Qing-jie CAO3,Hui-long LI1,Qi-xing HU1   

  1. 1.School of Microelectronics, Tianjin University, Tianjin 300072,China
    2.Tianjin Key Laboratory of Imaging and Sensing Microelectronic Technology, Tianjin 300072,China
    3.School of Mathematical Sciences, Tianjin Normal University, Tianjin 300387,China
  • Received:2019-06-03 Online:2020-09-01 Published:2020-09-16

摘要:

为提升在低辐射剂量条件下能谱式计算机断层扫描(CT)的重建图像质量,提出了一种基于复合感知损失函数的生成对抗网络去噪模型。此方法将像素空间与人类感知的特征空间结合到网络的生成对抗过程,并引入残差学习解决网络层数加深导致的图像细节丢失问题。通过采用多个能段的CT图像作为输入,同时利用了能段内的空间相关性和能段间的能量相关性,提高了能谱CT图像的视觉灵敏度。实验结果表明:该方法将峰值信噪比提高约5 dB,结构相似性指数提高约0.2,特征相似性指数提高约0.06。与当前的低剂量CT影像去噪算法相比,本文模型可实现更好的噪声去除效果,同时能保留诊断必要的细节信息,显著提高了低剂量能谱CT图像的质量。

关键词: 信息处理技术, 低剂量, 生成对抗网络, 残差学习

Abstract:

The continuous development and widespread use of computed tomography(CT) in medical practice have caused patients and doctors’ attention to the risk of cancer due to radiation dose. However, spectral CT can generate severe Poisson noise as a result of the division of multiple energy bins while reducing the radiation dose, which has an adverse effect on clinical diagnosis. Aiming at improving the reconstructed image quality of spectral CT under low radiation dose, a denoising model based on a generative adversarial network with the hybrid perceptual loss was proposed. The method combines pixel space with the human-perceived feature space into generative adversarial network, and introduces the residual learning to reuse features from different levels to avoid the loss of important details that come from the increase in the number of network layers. Then the Wasserstein distance is chosen instead of Jensen-Shannon to improve the stability of training. The inputs are from multiple energy bins, so that the spatial correlation of the reconstructed image from a single energy bin and the energy correlation between different reconstructed images from multiple energy bins can be utilized to improve the visual sensitivity of low dose spectral CT images. The experiments were performed on the simulated phantom data sets from Duke University Medical Center. The results show that the proposed method increases the PSNR by about 5 dB, the structure similarity (SSIM) value by about 0.2, and the feature similarity (FSIM) value about by 0.06. Compared with current advanced low-dose CT denoising method, the hybrid loss function based generative and adversarial network achieves a better noise removal, while retaining the necessary details for diagnosis. Moreover, the spectral information also plays a positive role in the recovery of noise images, and the use of hybrid perceptual loss is superior to any single loss. Overall, the proposed generative adversarial network with a hybrid loss method can significantly improve the quality of low-dose spectral CT images.

Key words: information processing technology, low dose, generative adversarial network, residual learning

中图分类号: 

  • TP391.4

图1

低剂量能谱CT去噪网络框架图"

图2

生成对抗网络内部结构图"

图3

图像①不同算法去噪效果对比图"

图4

图像②不同算法去噪效果对比图"

表1

图像①②的客观评价指标"

图像序号能量区间指标LDCTCNNWGAN-VGGHPWGAN-2DHPWGAN-3D
30~80 keVPSNR/dB32.62

36.94

(+4.32)

34.03

(+1.41)

35.54

(+2.92)

37.66

(+5.04)

SSIM0.7219

0.8888

(+0.1669)

0.8540

(+0.1321)

0.8683

(+0.1464)

0.8933

(+0.1714)

FSIM0.9098

0.9630

(+0.0532)

0.9427

(+0.0329)

0.9591

(+0.0493)

0.9744

(0.0646)

80~120 keVPSNR/dB33.15

37.94

(+4.79)

35.52

(+2.37)

36.79

(+3.64)

38.95

(+5.80)

SSIM0.7373

0.9158

(+0.1785)

0.9156

(+0.1783)

0.9157

(+0.1784)

0.9270

(+0.1897)

FSIM0.9132

0.9666

(+0.0534)

0.9500

(+0.0368)

0.9640

(+0.0508)

0.9793

(+0.0661)

30~80 keVPSNR/dB32.37

36.07

(+3.70)

33.25

(+0.88)

34.76

(+2.39)

37.09

(+4.72)

SSIM0.6929

0.8708

(+0.1779)

0.8209

(+0.1280)

0.8573

(+0.1644)

0.8810

(+0.1881)

FSIM0.9086

0.9604

(+0.0518)

0.9412

(+0.0326)

0.9565

(+0.0479)

0.9716

(+0.0630)

80~120 keVPSNR/dB32.88

37.76

(+4.88)

35.54

(+2.66)

36.82

(+3.94)

38.70

(+5.82)

SSIM0.7137

0.9085

(+0.1948)

0.9054

(+0.1917)

0.9070

(+0.1933)

0.9198

(+0.2061)

FSIM0.9134

0.9630

(+0.0496)

0.9486

(+0.0352)

0.9613

(+0.0479)

0.9753

(+0.0619)

图5

不同算法的图像评价指标在测试集上的统计平均值"

图6

不同损失函数去噪效果图"

表2

不同损失函数在测试集上的平均指标比较"

能量区间指标MSE 损失纹理 损失感知 损失复合感知损失
30~80 keVPSNR/dB37.3935.6534.8037.41
SSIM0.81690.84670.79120.8919
FSIM0.97120.96330.94850.9738
80~120 keVPSNR/dB39.2437.4136.6138.70
SSIM0.92360.91040.90880.9255
FSIM0.97380.96910.95590.9774
1 Brenner D J, Hall E J. Computed tomography-an increasing source of radiation exposure[J]. New England Journal of Medicine, 2007, 357(22): 2277-2284.
2 de Gonzalez A B, Mahesh M, Kim K P, et al. Projected cancer risks from computed tomographic scans performed in the united states in 2007[J]. Archives of Internal Medicine, 2009, 169(22): 2071-2077.
3 Wang J, Li T F, Lu H B, et al. Penalized weighted least-squares approach to sinogram noise reduction and image reconstruction for lowdose X-ray computed tomography[J]. IEEE Transactions on Medical Imaging, 2006, 25(10): 1272-1283.
4 Beister M, Kolditz D, Kalender W A. Iterative reconstruction methods in X-ray CT[J]. Physica Medica, 2012, 28(2): 94-108.
5 Liu Y, Ma J H, Fan Y, et al. Adaptive-weighted total variation minimization for sparse data toward low-dose X-ray computed tomography image reconstruction[J]. Physics in Medicine & Biology, 2012, 57(23): 7923-7956.
6 Yang C, Yin X D, Shi L Y, et al. Improving abdomen tumor low-dose CT images using a fast dictionary learning based processing[J]. Physics in Medicine & Biology, 2013, 58(16): 5803-5820.
7 Fumene F P, Vinegoni C, Gros J, et al. Block matching 3D random noise filtering for absorption optical projection tomography[J]. Physics in Medicine & Biology, 2010, 55(18): 5401-5415.
8 郜峰利, 陶敏, 李雪妍, 等. 基于深度学习的CT影像脑卒中精准分割[J]. 吉林大学学报: 工学版, 2020, 50(2): 678-684.
Gao Feng-li, Tao Min, Li Xue-yan, et al. Accurate segmentation of stroke in CT image based on deep learning[J]. Journal of Jilin University(Engineering and Technology Edition), 2020, 50(2): 678-684.
9 司建波, 杨芳, 郭蔚莹, 等. 基于BP神经网络的两阶段疾病预测模型[J]. 吉林大学学报: 工学版, 2013, 43(): 481-484.
Si Jian-bo, Yang Fang, Guo Wei-ying, et al. Two-stage disease prediction model based on BP neural network[J]. Journal of Jilin University(Engineering and Technology Edition), 2013, 43(Sup.1): 481-484.
10 Chen H, Zhang Y, Zhang W H, et al. Low-dose CT via convolutional neural network[J]. Biomedical Optics Express, 2017, 8(2): 679-694.
11 Wolterink J M, Leiner T, Viergever M A, et al. Generative adversarial networks for noise reduction in low-dose CT[J]. IEEE Transactions on Medical Imaging, 2017, 36(12): 2536-2545.
12 Yang Q S, Yan P K, Zhang Y B, et al. Low dose CT image denoising using a generative adversarial network with Wasserstein distance and perceptual loss[J]. IEEE Transactions on Medical Imaging, 2018, 37(6): 1348-1357.
13 Zhang Y B, Mou X Q, Wang G, et al. Tensor-based dictionary learning for spectral CT reconstruction[J]. IEEE Transactions on Medical Imaging, 2017, 36(1): 142-154.
14 Shi Z F, Yang H Y, Cong W X, et al. An edge-on charge-transfer design for energy-resolved X-ray detection[J]. Physics in Medicine and Biology, 2016, 61(11): 4183-4200.
[1] 张薇,韩勇,金铭,乔晓林. 基于托普利兹矩阵集重构的相干信源波达方向估计[J]. 吉林大学学报(工学版), 2020, 50(2): 703-710.
[2] 程艳芬,姚丽娟,袁巧,陈先桥. 水下视频图像清晰化方法[J]. 吉林大学学报(工学版), 2020, 50(2): 668-677.
[3] 于晓辉,张志成,李新波,孙晓东. 基于状态空间模型的指数衰减正弦信号参数估计[J]. 吉林大学学报(工学版), 2019, 49(6): 2083-2088.
[4] 刘富, 权美静, 王柯, 刘云, 康冰, 韩志武, 侯涛. 仿蝎子振源定位机理的位置指纹室内定位方法[J]. 吉林大学学报(工学版), 2019, 49(6): 2076-2082.
[5] 马子骥,卢浩,董艳茹. 双通道单图像超分辨率卷积神经网络[J]. 吉林大学学报(工学版), 2019, 49(6): 2089-2097.
[6] 郭继昌,吴洁,郭春乐,朱明辉. 基于残差连接卷积神经网络的图像超分辨率重构[J]. 吉林大学学报(工学版), 2019, 49(5): 1726-1734.
[7] 曹运合,曾丽,王宇. 基于特征空间的子阵级自适应和差波束测角方法[J]. 吉林大学学报(工学版), 2019, 49(5): 1735-1744.
[8] 卢洋,王世刚,赵文婷,赵岩. 基于离散Shearlet类别可分性测度的人脸表情识别方法[J]. 吉林大学学报(工学版), 2019, 49(5): 1715-1725.
[9] 董超,刘晶红,徐芳,王仁浩. 光学遥感图像舰船目标快速检测方法[J]. 吉林大学学报(工学版), 2019, 49(4): 1369-1376.
[10] 王柯俨,胡妍,王怀,李云松. 结合天空分割和超像素级暗通道的图像去雾算法[J]. 吉林大学学报(工学版), 2019, 49(4): 1377-1384.
[11] 托乎提努尔,张海龙,王杰,王娜,冶鑫晨,王万琼. 基于图形处理器的高速中值滤波算法[J]. 吉林大学学报(工学版), 2019, 49(3): 979-985.
[12] 付银娟,李勇,徐丽琴,张昆辉. NLFM⁃Costas射频隐身雷达信号设计及分析[J]. 吉林大学学报(工学版), 2019, 49(3): 994-999.
[13] 苏寒松,代志涛,刘高华,张倩芳. 结合吸收Markov链和流行排序的显著性区域检测[J]. 吉林大学学报(工学版), 2018, 48(6): 1887-1894.
[14] 应欢,刘松华,唐博文,韩丽芳,周亮. 基于自适应释放策略的低开销确定性重放方法[J]. 吉林大学学报(工学版), 2018, 48(6): 1917-1924.
[15] 陆智俊,钟超,吴敬玉. 星载合成孔径雷达图像小特征的准确分割方法[J]. 吉林大学学报(工学版), 2018, 48(6): 1925-1930.
Viewed
Full text


Abstract

Cited

  Shared   
  Discussed   
No Suggested Reading articles found!