Low-dose CT Reconstruction Based on Deep Energy Models
-
摘要: 降低计算机断层扫描(CT)的剂量对于降低临床应用中的辐射风险至关重要,深度学习的快速发展和广泛应用为低剂量CT成像算法的发展带来了新的方向。与大多数受益于手动设计的先验函数或有监督学习方案的现有先验驱动算法不同,本文使用基于深度能量模型来学习正常剂量CT的先验知识,然后在迭代重建阶段,将数据一致性作为条件项集成到低剂量CT的迭代生成模型中,通过郎之万动力学迭代更新训练的先验,实现低剂量CT重建。实验比较,证明所提方法的降噪和细节保留能力优良。Abstract: Reducing the dose of computed tomography (CT) is essential for reducing the radiation risk in clinical applications. With the rapid development and wide application of deep learning, it has brought new directions for the development of low-dose CT imaging algorithms. Unlike most existing prior-driven algorithms that benefit from manually designed prior functions or supervised learning schemes, in this paper, we use an energy-based deep model to learn the prior knowledge of normal-dose CT, and then in the iterative reconstruction phase, we integrate data consistency as a conditional item into the iterative generation model of low-dose CT, and realize the low-dose CT reconstruction through the prior experience of iterative updating training of Langevin dynamics. The experimental results show that the proposed method hold excellent noise reduction and detail retention capabilities.
-
Key words:
- low-dose CT /
- deep learning /
- Langevin dynamics /
- deep energy based model
-
表 1 EBM-LDCT算法的重建过程
Table 1. The reconstruction process of the EBM-LDCT algorithm
算法1 初始化:输入数据${\tilde x^0}\sim N(0,{\boldsymbol{I}})$,动量项${{\boldsymbol{w}}^0}$,参数$ \beta $,松弛因子$ \gamma $,迭代次数$ T $
开始循环:$t = 1,{\text{ }}2,{\text{ }} \cdots ,{\text{ }}T$
设定$ { {\boldsymbol{z} }^t}\sim N(0,{\boldsymbol{I} }) $
根据公式(10)更新:$ {\tilde {\boldsymbol{u} }^t} = {\tilde {\boldsymbol{x} }^{t - 1} } - \displaystyle\dfrac{ {\,\varepsilon \,} }{2}{\nabla _x}E{}_\theta \left( { { {\tilde {\boldsymbol{x} } }^{t - 1} } } \right) + \sqrt \varepsilon { {\boldsymbol{z} }^t} $
根据公式(8)更新:$ {\tilde {\boldsymbol{x} }^t} = { {\boldsymbol{w} }^{t - 1} } - \displaystyle\dfrac{ { {\kern 1 pt} { {\boldsymbol{A} }^{\rm{T} } }\left( { {\boldsymbol{A} }{ {\tilde {\boldsymbol{x} } }^{t - 1} } - {\boldsymbol{y} } } \right) + \beta \left( { { {\tilde {\boldsymbol{x} } }^{t - 1} } - { {\tilde {\boldsymbol{u} } }^t} } \right){\kern 1 pt} } }{ { { {\boldsymbol{A} }^{\rm{T} } }{\boldsymbol{A} }{\boldsymbol{1} } + \beta {\boldsymbol{1} } } } $
最后更新:${{\boldsymbol{w}}^t} = {\tilde {\boldsymbol{x}}^t} + \gamma \left( { { {\tilde {\boldsymbol{x}}}^t} - { {\tilde {\boldsymbol{x}}}^{t - 1} } } \right)$
${\tilde {\boldsymbol{x}}^t} \leftarrow {{\boldsymbol{w}}^t}$
结束表 2 使用CIRS数据集的多种低剂量CT重建方法定量结果比较
Table 2. Comparison of quantitative results of multiple LDCT reconstruction methods using CIRS dataset
方法 MAE PSNR/dB SSIM/10-2 FBP(Ramp-filter) 9.48±0.64 40.67±0.59 94.36±0.87 TV 7.64±0.41 42.12±0.35 96.58±0.51 K-SVD 7.01±0.19 42.86±0.34 97.01±0.36 RED-CNN 6.74±0.07 41.76±0.12 97.47±0.16 EBM-LDCT 6.29±0.23 42.73±0.20 97.81±0.27 表 3 使用AAPM数据集的多种LDCT重建方法定量结果比较
Table 3. Comparison of quantitative results of multiple LDCT reconstruction methods using AAPM dataset
方法 MAE PSNR/dB SSIM/10-2 FBP(Ramp-filter) 67.69±13.16 28.68±2.03 44.43±7.18 TV 29.65±4.60 34.98±1.32 86.10±3.48 K-SVD 21.00±1.04 35.68±2.26 81.98±2.41 RED-CNN 16.97±1.02 39.37±1.87 94.78±0.56 EBM-LDCT 18.75±1.31 39.56±0.54 94.16±0.68 -
[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. doi: 10.1056/NEJMra072149 [2] KALRA M K, MAHER M M, TOTH T L, et al. Strategies for CT radiation dose optimization[J]. Radiology, 2004, 230(3): 619−628. doi: 10.1148/radiol.2303021726 [3] 韩泽芳, 上官宏, 张雄, 等. 基于深度学习的低剂量CT成像算法研究进展[J]. CT理论与应用研究, 2022, 31(1): 118-136. DOI: 10.15953/j.1004-4140.2022.31.01.14.HAN Z F, SHANGGUAN H, ZHANG X, et al. Research progress of low-dose CT imaging algorithm based on deep learning[J]. CT Theory and Applications, 2022, 31(1): 118-136. DOI: 10.15953/j.1004-4140.2022.31.01.14. (in Chinese). [4] TIAN Z, JIA X, YUAN K, et al. Low-dose CT reconstruction via edge-preserving total variation regularization[J]. Physics in Medicine & Biology, 2011, 56(18): 5949. [5] MA J, LIANG Z, FAN Y, et al. A study on CT sinogram statistical distribution by information divergence theory[C]//2011 IEEE Nuclear Science Symposium Conference Record. IEEE, 2011: 3191-3196. [6] ELBAKRI I A, FESSLER J A. Statistical image reconstruction for polyenergetic X-ray computed tomography[J]. IEEE Transactions on Medical Imaging, 2002, 21(2): 89−99. doi: 10.1109/42.993128 [7] ZHANG K, ZUO W, CHEN Y, et al. Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising[J]. IEEE Transactions on Image Processing, 2017, 26(7): 3142−3155. doi: 10.1109/TIP.2017.2662206 [8] LEFKIMMIATIS S. Universal denoising networks: A novel CNN architecture for image denoising[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018: 3204-3213. [9] ZHANG K, ZUO W, ZHANG L. FFDNet: Toward a fast and flexible solution for CNN-based image denoising[J]. IEEE Transactions on Image Processing, 2018, 27(9): 4608−4622. doi: 10.1109/TIP.2018.2839891 [10] CHEN H, ZHANG Y, ZHANG W, et al. Low-dose CT via convolutional neural network[J]. Biomedical Optics Express, 2017, 8(2): 679−694. doi: 10.1364/BOE.8.000679 [11] CHEN H, ZHANG Y, KALRA M K, et al. Low-dose CT with a residual encoder-decoder convolutional neural network[J]. IEEE Transactions on Medical Imaging, 2017, 36(12): 2524−2535. doi: 10.1109/TMI.2017.2715284 [12] WÜRFL T, GHESU F C, CHRISTLEIN V, et al. Deep learning computed tomography[C]//International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, Cham, 2016: 432-440. [13] CHENG L, AHN S, ROSS S G, et al. Accelerated iterative image reconstruction using a deep learning based leapfrogging strategy[C]//International Conference on Fully Three-Dimensional Image Reconstruction in Radiology and Nuclear Medicine, 2017: 715-720. [14] HAN Y S, YOO J, YE J C. Deep residual learning for compressed sensing CT reconstruction via persistent homology analysis[J]. arXiv preprint arXiv: 1611.06391, 2016. [15] JIN K H, MCCANN M T, FROUSTEY E, et al. Deep convolutional neural network for inverse problems in imaging[J]. IEEE Transactions on Image Processing, 2017, 26(9): 4509−4522. doi: 10.1109/TIP.2017.2713099 [16] RONNEBERGER O, FISCHER P, BROX T. U-net: Convolutional networks for biomedical image segmentation[C]//International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, Cham, 2015: 234-241. [17] HE K, ZHANG X, REN S, et al. Deep residual learning for image recognition[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016: 770-778. [18] BAGUER D O, LEUSCHNER J, SCHMIDT M. Computed tomography reconstruction using deep image prior and learned reconstruction methods[J]. Inverse Problems, 2020, 36(9): 094004. doi: 10.1088/1361-6420/aba415 [19] KINGMA D P, WELLING M. Auto-encoding variational bayes[J]. arXiv preprint arXiv: 1312.6114, 2013. [20] DINH L, KRUEGER D, BENGIO Y. Nice: Non-linear independent components estimation[J]. arXiv preprint arXiv: 1410.8516, 2014. [21] GOODFELLOW I, POUGET-ABADIE J, MIRZA M, et al. Generative adversarial nets[J]. Advances in Neural Information Processing Systems, 2014: 27. [22] NOWOZIN S, CSEKE B, TOMIOKA R. F-GAN: Training generative neural samplers using variational divergence minimization[C]//Proceedings of the 30 th International Conference on Neural Information Processing Systems, 2016: 271-279. [23] ARJOVSKY M, CHINTALA S, BOTTOU L. Wasserstein generative adversarial networks[C]// International Conference on Machine Learning. PMLR, 2017: 214-223. [24] 蔡宁, 王世杰, 陈璐杰, 等. 基于渐进式网络处理的低剂量Micro-CT成像方法[J]. CT理论与应用研究, 2020,29(4): 435−446. DOI: 10.15953/j.1004-4140.2020.29.04.06.CAI N, WANG S J, CHEN L J. Low-dose Micro-CT imaging method based on progressive network processing[J]. CT Theory and Applications, 2020, 29(4): 435−446. DOI: 10.15953/j.1004-4140.2020.29.04.06. (in Chinese). [25] NIU C, LI M Z, FAN F L, et al. Suppression of correlated noises with similarity-based unsupervised deep learning[J]. arXiv: 2011.03384. [26] WU D, KIM K, E L FAKHRI G, et al. Iterative low-dose CT reconstruction with priors trained by artificial neural network[J]. IEEE Transactions on Medical Imaging, 2017, 36(12): 2479−2486. doi: 10.1109/TMI.2017.2753138 [27] ZHANG F Q, ZHANG M H, QIN B J, et al. REDAEP: Robust and enhanced denoising autoencoding prior for sparse-view CT reconstruction[J]. IEEE Transactions on Radiation and Plasma Medical Sciences, 2021, 5(1): 108−119. doi: 10.1109/TRPMS.2020.2989634 [28] LECUN Y, CHOPRA S, HADSELL R, et al. A tutorial on energy-based learning[M]. Predicting Structured Data, 2006: 191-241. [29] DU Y, MORDATCH I. Implicit generation and modeling with energy based models[J]. Advances in Neural Information Processing Systems, 2019: 32. [30] NIJKAMP E, HILL M, ZHU S C, et al. Learning non-convergent non-persistent short-run MCMC toward energy-based model[J]. arXiv preprint arXiv: 1904.09770, 2019. [31] WELLING M, TEH Y W, Bayesian learning via stochastic gradient Langevin dynamics[C]// International Conference on International Conference on Machine Learning. Omnipress, 2011. [32] ERDOGAN H, FESSLER J A. Monotonic algorithms for transmission tomography[C]//5th IEEE EMBS International Summer School on Biomedical Imaging, 2002. [33] DEVOLDER O, GLINEUR F, NESTEROV Y. First-order methods of smooth convex optimization with inexact oracle[J]. Mathematical Programming, 2014, 146(1): 37−75. [34] YIN X, ZHAO Q, LIU J, et al. Domain progressive 3D residual convolution network to improve low-dose CT imaging[J]. IEEE transactions on medical imaging, 2019, 38(12): 2903−2913. doi: 10.1109/TMI.2019.2917258 -
2021.077-资源附件文件.doc
-