|
1.INTRODUCTIONComparing with the conventional computed tomography (CT), photon-counting CT (PCCT) can obtain multiple measurements of the scanned object at multi-energy bins and provide abundant energy-dependent material-specific information. Due to the energy discrimination capability, PCCT can effectively improve contrast-to-noise ratio, increase the dose efficiency and reduce electronic noise.1, 2 However, the collected PCCT measurements in the narrow energy bins are corrupted by serious quantum noise,3 and the PCCT image quality degrades obviously due to the limited photons. To solve this problem, many statistical iteration reconstruction (SIR) methods have been proposed in the past decades. The main idea of the SIR is to construct a reconstruction model with data fidelity and regularization terms, where the first term incorporates the statistical property of X-ray photons and the second term provides the prior information of the desired PCCT images. For example, Rigie et al. introduced a total nuclear variation regularization to leverage similar gradient information and improve the image quality.4 Kim et al. developed a patch-based low-rank regularization to maintain the image structures and reduce noise.5 Semerci et al. combined a tensor nuclear norm and a total variation regularization to suppress noise.6 Zhang et al. proposed to deliver inner spectrum correlation information and constructed a tensor-based dictionary learning strategy.7 Niu et al. considered the self-similarity of the spectral CT images and proposed a non-local low-rank and sparse matrix decomposition method.3 Wu et al. proposed to encourage the similarity of spectral CT images by utilizing a cube-based tensor regularization.8 Recently, Zeng et al. proposed a full-spectrum-knowledge-aware tensor by imposing the global correlation, piecewise smooth and latent full-spectrum properties of PCCT images.9 These methods have been shown great potential in preserving image details and suppressing noise. However, there remain some challenges in practice. First, the SIR methods usually utilize fan-beam geometry and the computation costs will be a burden for cone-beam geometry. Second, SIR methods are sensitive for the hyper-parameters, and appropriate parameters selection is needed for different clinical applications. Recently, deep learning (DL) technology has been widely adopted in CT imaging field. In spectral CT imaging field, Lu et al. utilized DL-based method for material decomposition.10 Fang et al. proposed to remove ring artifacts for PCCT data by using DL-based method.11 Wu et al. employed a DL-based method for reconstructing PCCT images.12 It is shown that the DL-based methods achieved competing results compared with the SIR methods. However, the current DL-based methods need a large quantity of paired data (i.e., noisy and high-quality data) to obtain a desired model by supervised training strategy. Moreover, collecting large-scale spectral CT data is time-consuming, and the PCCT data in clinics is hard to be obtained. By the way, a number of energy-integrating detector (EID) data, which is easily to be obtained, is not yet included in training the DL-based method for PCCT imaging. Therefore, we present an unsupervised DL-based method in the image domain by utilizing the prior information of the paired EID dataset (i.e., low-dose images/high-quality ones). Specifically, we first initialize supervised and unsupervised networks for EID and PCCT images, respectively. Then, the supervised network is trained on well paired EID dataset and serves as the prior knowledge to regularize the unsupervised PCCT network training. Moreover, a data-fidelity term for characterizing the PCCT image characteristics is constructed as the self-supervised loss. Finally, with the prior knowledge and self-supervised terms, we can train the network for PCCT images following an unsupervised learning strategy. We call the presented DL-based model as full-spectrum-knowledge-aware DL-based network, shorten as “FSANet”. We evaluate the presented FSANet and other reconstruction methods on synthesized clinical data. Experimental results demonstrate that the presented FSANet outperforms the competing methods in terms of qualitative and quantitative metrics. 2.METHODSConsidering the spatial and energy dimensions of the spectral data, the PCCT imaging model can be expressed as follows: where y = {yn, n ≤ N} and X* = The images reconstructed from Y suffer from noise and artifacts. To improve image quality, DL-based methods are feeded with the PCCT images and produce the denoised ones. It can be expressed as follows: where where L is the user defined loss function and Xtarget are the FBP-reconstructed images from the high-dose measurements. In order to utilize the prior information of the network pre-trained on paired EID dataset, we present a full-spectrum-knowledge-aware (FSA) loss function, as follows: where In summary, the total loss function of the presented model is expressed as follows: where α is a hyper-parameter of the loss function. It can be seen that the presented model engages an unsupervised learning strategy where only noisy PCCT images is involved in the training stage. Fig. 1 illustrates the presented FSANet method for PCCT image recovery. Finally, we utilize Adam14 to optimize the parameters in the network. Figure 1.Illustration of the presented FSANet method for PCCT image recovery. The pipeline with black arrows denote the data flow of the PCCT images. The pipeline with orange arrows denote calculations of loss functions. The supervised and unsupervised networks have the same architecture. It should be noted that the optimization for the parameters of the unsupervised network follows an unsupervised strategy, and only noisy PCCT images are involved during its training period. ![]() 3.RESULTS3.1Implementation DetailsThe network In experiments, we simulate 3000 cases to establish the whole dataset, and randomly select 2000, 500 and 500 cases for training, validation and testing datasets, respectively. In training period, the learning rate, batch size and training epoch are 1e−4, 6 and 2000, respectively. A modified residual network17 is selected as the backbone network for the supervised and unsupervised networks of the presented FSANet method. Both networks are implemented in Python with PyTorch package on a NVIDIA Tesla K40c GPU. 3.2Qualitative AnalysisFig. 2 illustrates the visual comparisons of the presented and compared methods on Case 1. The images at normal dose are chosen as the ground truth. It can be observed that FBP algorithm suffers from noise. TDL effectively removes the noise-induced artifacts and improves the image quality, but losses the image resolution. On the contrary, the presented FSANet produces more remarkable results closing to the Supervised Net and the ground truth in terms of noise reduction and structure preservation. Moreover, zoomed in regions-of-interest (ROIs) by the blue boxes in Fig. 2 is selected for better visual inspection. It can be seen that TDL smooths the structure edges, but the Supervised Net and presented FSANet methods maintain the image details. Figure 2.Results of the presented and compared methods on Case 1. The display windows from Bin 1 to 5 are [0.005, 0.008], [0.0028, 0.005], [0.002, 0.004], [0.0015, 0.0035] and [0.0015,0.0028] mm−1, respectively. Zoomed ROIs indicated by the blue boxes are displayed for better visualization. ![]() Fig. 3 shows the results of different methods on Case 2. Similar to Case 1, FBP induces noise to the images, TDL is prone to produce blurry results, and the presented FSANet avoids over smoothing and preserves structure details. The zoomed in ROIs indicated by the red boxes in Fig. 3 also demonstrate the advanced performance of the presented FSANet method. Fig. 4(a) and (b) show the profiles indicated by the green and orange lines in the Fig. 2 and Fig. 3, respectively. From the results, we can observe that the presented FSANet produces the closet results to the ground truth compared with the FBP and TDL methods. 3.3Quantitative AnalysisIn this study, peak signal-to-noise (PSNR) and root-mean-square-error (RMSE) are utilized to quantify the performances of different methods. Tab. 1 lists the quantitative measurements of the results from different methods on the whole testing dataset. From the results, it can be seen that the presented FSANet achieves better results among all metrics, compared with FBP and TDL methods. Therefore, the qualitative and quantitative results demonstrate that the presented FSANet method achieves superior results to the FBP and TDL methods. Table 1.Quantitative measurements on the reconstruction results on the testing dataset from the different methods.
4.DISCUSSION AND CONCLUSIONDL-based methods have been shown promising performance in conventional CT imaging, and it has also inspired the application in PCCT imaging. However, most of them are supervise-based and need a large quantity of paired training dataset, which is hard to be obtained for PCCT. To address this intrinsic limitation, in this work, we presented a DL-based PCCT denoising method with an unsupervised learning strategy, called “FSANet”. Specifically, we first trained a denoising network on paired EID dataset and served it as a prior for PCCT images. Then, we constructed this prior network as a training loss to regularize the PCCT network learning in an unsupervised manner. Moreover, a self-supervised loss of the noisy PCCT images is introduced to promote data-fidelity of the PCCT images. Finally, with the mentioned two loss terms, we obtained the presented FSANet method. Simulation experiments demonstrated the feasibility and effectiveness of the presented FSANet method. In the future, clinical patient studies would be involved to further demonstrate the denosing performance of the presented FSANet method. ACKNOWLEDGMENTSThis work was supported in part by the NSFC under Grant U21A6005 and Grant U1708261, the National Key R&D Program of China under Grant No. 2020YFA0712200, and Young Talent Support Project of Guangzhou Association for Science and Technology. REFERENCESM. J. Willemink, M. Persson, A. Pourmorteza, N. J. Pelc, and D. Fleischmann,
“Photon-counting CT: Technical principles and clinical prospects,”
Radiology, 289
(2), 293
–312
(2018). https://doi.org/10.1148/radiol.2018172656 Google Scholar
K. Taguchi, J. S. Iwanczyk,
“Vision 20/20: Single photon counting x-ray detectors in medical imaging,”
Medical Physics, 40
(10), 100901
(2013). https://doi.org/10.1118/1.4820371 Google Scholar
S. Niu, G. Yu, J. Ma, and J. Wang,
“Nonlocal low-rank and sparse matrix decomposition for spectral CT reconstruction,”
Inverse Problems, 34
(2), 024003
(2018). https://doi.org/10.1088/1361-6420/aa942c Google Scholar
D. S. Rigie and P. J. La Rivire,
“Joint reconstruction of multi-channel, spectral CT data via constrained total nuclear variation minimization,”
Physical in Medical & Biology, 60
(5), 1741
–1762
(2015). https://doi.org/10.1088/0031-9155/60/5/1741 Google Scholar
K. Kim and et al,
“Sparse-view spectral CT reconstruction using spectral patch-based low-rank penalty,”
IEEE Transactions on Medical Imaging, 34
(3), 748
–760
(2015). https://doi.org/10.1109/TMI.2014.2380993 Google Scholar
O. Semerci, N. Hao, M. E. Kilmer, and E. L. Miller,
“Tensor-based formulation and nuclear norm regularization for multienergy computed tomography,”
IEEE Transactions on Image Processing, 23
(4), 1678
–1693
(2014). https://doi.org/10.1109/TIP.83 Google Scholar
Y. Zhang, X. Mou, G. Wang, and H. Yu,
“Tensor-based dictionary learning for spectral CT reconstruction,”
IEEE Transactions on Medical Imaging, 36
(1), 142
–154
(2017). https://doi.org/10.1109/TMI.2016.2600249 Google Scholar
W. Wu, F. Liu, Y. Zhang, Q. Wang, and H. Yu,
“Non-local low-rank cube-based tensor factorization for spectral CT reconstruction,”
IEEE Transactions on Medical Imaging, 38
(4), 1079
–1093
(2019). https://doi.org/10.1109/TMI.42 Google Scholar
D. Zeng, L. Yao, Y. Ge, S. Li, Q. Xie, H. Zhang, and et al,
“Full-spectrum-knowledge-aware tensor model for energy-resolved CT iterative reconstruction,”
IEEE Transactions on Medical Imaging, 39 2831
–2843
(2020). https://doi.org/10.1109/TMI.42 Google Scholar
Y. Lu, M. Kowarschik, X. Huang, Y. Xia, J. Choi, S. Chen, and et al,
“A learning-based material decomposition pipeline for multi-energy X-ray imaging,”
Medical physics, 46 689
–703
(2019). https://doi.org/10.1002/mp.2019.46.issue-2 Google Scholar
W. Fang, L. Li, and Z. Chen,
“Removing ring artefacts for photon-counting detectors using neural networks in different domains,”
IEEE Access, 8 42447
–42457
(2020). https://doi.org/10.1109/Access.6287639 Google Scholar
W. Wu, D. Hu, C. Niu, L. Broeke, A. Butler, P. Cao, and et al,
“Deep learning based spectral CT imaging,”
Neural Networks, 144 342
–358
(2021). https://doi.org/10.1016/j.neunet.2021.08.026 Google Scholar
G. L. Zeng,
“Medical image reconstruction: a conceptual tutorial,”
Springer,2010). Google Scholar
D. P. Kingma and J. Ba,
“Adam: A method for stochastic optimization,”
arXiv preprint, arXiv: 1412.6980,
(2014). Google Scholar
M. Arjovsky, S. Chintala, and L. Bottou,
“Wasserstein GAN,”
https://arxiv.org/abs/1701.07875 Google Scholar
J. Punnoose, J. Xu, A. Sisniega, W. Zbijewski, and J. H. Siewerdsen,
“Technical note: Spektr 3.0-A computational tool for X-ray spectrum modeling and analysis,”
Medical Physics, 43
(8), 4711
–4717
(2016). https://doi.org/10.1118/1.4955438 Google Scholar
K. He, X. Zhang, S. Ren, et al.,
“Deep residual learning for image recognition,”
in Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR),
770
–778
(2016). Google Scholar
|