基于迭代残差网络的双能CT图像材料分解研究

王冲旭; 陈平; 潘晋孝; 刘宾

doi:10.15953/j.1004-4140.2022.31.01.05

基于迭代残差网络的双能CT图像材料分解研究

王冲旭^{1a, 2,},
陈平^{1b, 2, ,},
潘晋孝^{1a, 1b, 2},
刘宾^{1b, 2}

1.
理学院；中北大学
2.
信息与通信工程学院, 太原 030051 中北大学
2.
中北大学信息探测与处理山西省重点实验室, 太原 030051

基金项目: 国家自然科学基金（面向金属基复合材料微结构表征的X射线多谱CT成像方法研究（61801437）；基于深度学习的递变能量多谱CT成像表征方法研究（61871351）；基于深度学习的低剂量CT重建与影像识别（61971381））。

详细信息

作者简介:
王冲旭: 男，中北大学硕士研究生，主要从事X射线成像、双能材料分解方面的研究，E-mail：1162391239@qq.com

陈平: 男，中北大信息与通信工程学院教授，主要从事X射线成像、光电检测等方面的研究，E-mail：pc0912@163.com

中图分类号: TP 391
计量
- 文章访问数: 969
- HTML全文浏览量: 351
- PDF下载量: 130
出版历程
- 收稿日期: 2021-07-25
- 网络出版日期: 2021-11-05
- 刊出日期: 2022-01-31

Research on Material Decomposition of Dual-energy CT Image Based on Iterative Residual Network

WANG Chongxu^{1a, 2,},
CHEN Ping^{1b, 2, ,},
PAN Jinxiao^{1a, 1b, 2},
LIU Bin^{1b, 2}

1.
Taiyuan 030051, China School of Science
2.
Taiyuan 030051, China College of Information and Communication Engineering, North University of China
2.
Shanxi Key Laboratory of Signal Capturing and Processing, Taiyuan 030051, China

摘要

摘要:
双能计算机断层成像技术（DECT）由于其材料分解能力，在高级成像应用中发挥着重要作用。图像域分解直接对CT图像进行线性矩阵反演，但分解后的材料图像会受到噪声和伪影的严重影响。虽然各种正则化方法被提出来解决这个问题，但它们仍然面临着两个挑战：繁琐的参数调整和过度平滑导致的图像细节损失。为此，本文提出一种基于迭代残差网络的双能CT图像材料分解算法，直接求逆作为初始基图像，利用堆叠的双通道卷积神经网络替换迭代分解模型中的正则化项，构成深度迭代分解网络，该方法同时实现了材料分解和噪声抑制。实验结果表明，本文提出的迭代残差网络优于其他对比方法，能够在保持基图像边缘细节信息的同时有效抑制噪声和伪影。
- 计算机断层成像 /
- 双能 CT /
- 残差网络 /
- 噪声抑制
Abstract:
Dual energy computed tomography (DECT) plays an important role in the application of advanced imaging due to its material decomposition capability. Image domain decomposition can directly invert CT images through by linear matrix, but the decomposed material images will be seriously affected by noise and artifacts. Although various regularization methods have been proposed to solve this problem, they still face two challenges: tedious parameter adjustment and the loss of image details resulted from over-smoothing. Therefore, in this paper we proposes a dual energy CT image material decomposition algorithm based on iterative residual network. Direct inversion is used as the initial base image, and a stacking two-channel convolutional neural network is used to replace the regularization items in the iterative decomposition model to form a deep iterative decomposition network. This method can realize material decomposition and noise suppression simultaneously. Experimental results show that the iterative residual network proposed in this paper is superior to other comparison methods and can effectively suppress noise and artifacts while maintaining the edge details of the base image.
- computed tomography /
- dual-energy CT /
- residual network /
- noise suppression

HTML全文

图 1 IR-Net网络模型框架

Figure 1. Overall structure of IR-Net

下载: 全尺寸图片幻灯片

图 2 网络数据集的示例

（a）和（b）分别显示高能量和低能量下的重建图像；（c）和（d）显示测试切片的真实的水和骨的材料密度图像。

Figure 2. Example of the network dataset

下载: 全尺寸图片幻灯片

图 3 DIMD、DECT-ST、DECT-MULTRA和IR-Net的材料分解结果

Figure 3. Basis material images decomposed by DIMD, DECT-ST, DECT-MULTRA, and IR-Net

下载: 全尺寸图片幻灯片

图 4 DIMD、DECT-ST、DECT-EP和IR-Net与标签之前差值的绝对值

Figure 4. Images of absolute value of difference between ground truths and decomposed material images by DIMD, DECT-ST, DECT-MULTRA, and IR-Net

下载: 全尺寸图片幻灯片

图 5 DIMD、DECT-ST、DECT-MULTRA和IR-Net的材料分解结果

Figure 5. Basis material images decomposed by DIMD, DECT-ST, DECT-MULTRA, and IR-Net

下载: 全尺寸图片幻灯片

1 不同算法的定量评估结果

Materia	Index	DIMD	DECT-ST	DECT-MULTRA	IR-Net
	RMSE	0.0815	0.0500	0.0341	0.0169
water	PSNR	21.7735	26.0188	29.3327	35.4356
	SSIM	0.4201	0.9276	0.9112	0.9937
	RMSE	0.0786	0.0554	0.0440	0.0150
bone	PSNR	22.0949	25.1253	27.1373	36.4688
	SSIM	0.3399	0.8148	0.8949	0.9995

下载: 导出CSV

表 1 不同算法的定量评估结果

Table 1 Quantitative evaluation results of different algorithms

Materia	Index	DIMD	DECT-ST	DECT-MULTRA	IR-Net
	RMSE	0.0815	0.0500	0.0341	0.0169
water	PSNR	21.7735	26.0188	29.3327	35.4356
	SSIM	0.4201	0.9276	0.9112	0.9937
	RMSE	0.0786	0.0554	0.0440	0.0150
bone	PSNR	22.0949	25.1253	27.1373	36.4688
	SSIM	0.3399	0.8148	0.8949	0.9995

下载: 导出CSV

参考文献(15)

[1]	JOHNSON T R, KRAU B, SEDLMAIR M, et al. Material differentiation by dual energy CT: Initial experience[J]. European Radiology, 2007, 17(6): 1510-1517.
[2]	王文杰, 陈平, 潘晋孝, 等. 基于参考组分的双能CT成像方法[J]. CT理论与应用研究, 2021,30(1): 61−69. DOI: 10.15953/j.1004-4140.2021.30.01.06. WANG W J, CHEN P, PAN J X, et al. Dual-energy CT imaging method based on reference components[J]. CT Theory and Applications, 2021, 30(1): 61−69. DOI: 10.15953/j.1004-4140.2021.30.01.06. (in Chinese).
[3]	王李磊. 双能CT 图像重建算法研究[D]. 郑州: 解放军信息工程大学, 2016. WANG L L. Research on dual energy CT image reconstruction algorithm[D]. Zhengzhou: PLA University of information engineering, 2016. (in Chinese).
[4]	XUE Y, RUAN R S, HU X H, et al. Statistical image-domain multimaterial decomposition for dual-energy CT[J]. Medical Physics, 2017, 44(3): 886−901. doi: 10.1002/mp.12096
[5]	XUE D, NIU T Y, ZHU L. Combined iterative reconstruction and image-domain decomposition for dual energy CT using total-variation regularization[J]. Medical Physics, 2014, 41(5): 051909. DOI: 10.1118/1.4870375.
[6]	DING Q Q, NIU T Y, ZHANG X Q, et al. Image-domain multi-material decomposition for dual-energy CT based on prior information of material images[J]. Medical Physics, 2018: 45. DOI: 10.1002/mp.13001.
[7]	LI Z, RAVISHANKAR S, YONG L, et al. Image-domain material decomposition using data-driven sparsity models for dual-energy CT[C]//2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018). IEEE, 2018.
[8]	LI Z P, RAVISHANKAR S, LONG Y, et al. DECT-MULTRA: Dual-energy CT image decomposition with learned mixed material models and efficient clustering[J]. IEEE Transactions on Medical Imaging, 2020, 39(4): 1223−1234. DOI: 10.1109/TMI.2019.2946177.
[9]	ZHANG J L, GU Y B, TANG H, et al. Compressed sensing MR image reconstruction via a deep frequency-division network[J]. Neurocomputing, 2020, 384: 346−355. doi: 10.1016/j.neucom.2019.12.011
[10]	CHEN H, ZHANG Y, KALRA M K, et al. Low-dose CT with a residual encoder-decoder convolutional neural network (RED-CNN)[J]. IEEE Transactions on Medical Imaging, 2017, 36(99): 2524−2535.
[11]	SU T, SUN X, ZHANG Y, et al. DIRECT-net: A unified mutual-domain material decomposition network for quantitative dual-energy CT imaging[J]. 2020.
[12]	CHEN H, ZHANG Y, CHEN Y, et al. LEARN: Learned experts' assessment-based reconstruction network for sparse-data CT[J]. IEEE Transactions on Medical Imaging, 2018, (99): 1−1.
[13]	YANG Y, SUN J, LI H, et al. ADMM-CSNet: A deep learning approach for image compressive sensing[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020: 521−538.
[14]	ADLER J, OKTEM O. Learned primal-dual reconstruction[J]. IEEE Transactions on Medical Imaging, 2017: 1322−1332.
[15]	XU Y F, YAN B, ZHANG J F, et al. Image decomposition algorithm for dual-energy computed tomography via fully convolutional network[J]. Computational & Mathematical Methods in Medicine, 2018: 2527516. DOI: 10.1155/2018/2527516.