Department of Radiation Convergence Engineering, Yonsei University, 1 Yonseidae-gil, Wonju, 26493, Republic of Korea.
Department of Radiological Science, Yonsei University, 1 Yonseidae-gil, Wonju, 26493, Republic of Korea.
J Digit Imaging. 2021 Dec;34(6):1359-1375. doi: 10.1007/s10278-021-00467-w. Epub 2021 Sep 29.
Spectral computed tomography (CT) based on a photon-counting detector (PCD) is a promising technique with the potential to improve lesion detection, tissue characterization, and material decomposition. PCD-based scanners have several technical issues including operation in the step-and-scan mode and long data acquisition time. One straightforward solution to these issues is to reduce the number of projection views. However, if the projection data are under-sampled or noisy, it would be challenging to produce a correct solution without precise prior information. Recently, deep-learning approaches have demonstrated impressive performance for under-sampled CT reconstruction. In this work, the authors present a multilevel wavelet convolutional neural network (MWCNN) to address the limitations of PCD-based scanners. Data properties of the proposed method in under-sampled spectral CT are analyzed with respect to the proposed deep-running-network-based image reconstruction using two measures: sampling density and data incoherence. This work presents the proposed method and four different methods to restore sparse sampling. We investigate and compare these methods through a simulation and real experiments. In addition, data properties are quantitatively analyzed and compared for the effect of sparse sampling on the image quality. Our results indicate that both sampling density and data incoherence affect the image quality in the studied methods. Among the different methods, the proposed MWCNN shows promising results. Our method shows the highest performance in terms of various evaluation parameters such as the structural similarity, root mean square error, and resolution. Based on the results of imaging and quantitative evaluation, this study confirms that the proposed deep-running network structure shows excellent image reconstruction in sparse-view PCD-based CT. These results demonstrate the feasibility of sparse-view PCD-based CT using the MWCNN. The advantage of sparse view CT is that it can significantly reduce the radiation dose and obtain images with several energy bands by fusing PCDs. These results indicate that the MWCNN possesses great potential for sparse-view PCD-based CT.
基于光子计数探测器（PCD）的光谱 CT 是一种很有前途的技术，具有提高病灶检测、组织特征化和物质分解的潜力。基于 PCD 的扫描仪存在一些技术问题，包括在步进扫描模式下运行和数据采集时间长。解决这些问题的一个直接方法是减少投影视图的数量。然而，如果投影数据存在欠采样或噪声，没有精确的先验信息，就很难得到正确的解。最近，深度学习方法在欠采样 CT 重建方面表现出了令人印象深刻的性能。在这项工作中，作者提出了一种多级小波卷积神经网络（MWCNN）来解决基于 PCD 的扫描仪的限制。针对基于深度运行网络的图像重建，提出了一种新的方法，使用两种度量方法：采样密度和数据不连贯性，对所提出的方法在欠采样光谱 CT 中的数据特性进行了分析。这项工作提出了恢复稀疏采样的方法和四种不同的方法。通过模拟和真实实验，我们研究和比较了这些方法。此外，还对稀疏采样对图像质量的影响进行了定量分析和比较。我们的结果表明，在研究的方法中，采样密度和数据不连贯性都会影响图像质量。在不同的方法中，所提出的 MWCNN 显示出了有希望的结果。在各种评估参数方面，如结构相似性、均方根误差和分辨率，我们的方法表现出最高的性能。基于成像和定量评估的结果，本研究证实了所提出的深度运行网络结构在稀疏视图 PCD 基于 CT 中的出色图像重建能力。这些结果表明，MWCNN 具有在稀疏视图 PCD 基于 CT 中应用的可行性。稀疏视图 CT 的优势在于它可以显著减少辐射剂量，并通过融合 PCD 获得多个能谱带的图像。这些结果表明，MWCNN 具有在稀疏视图 PCD 基于 CT 中的巨大潜力。
