史再峰 ,李慧龙,程 明 ,曹清洁 ,王子菊
AuthorsHTML:史再峰 ^{1, 2}，李慧龙 ¹ ，程 明 ¹ ，曹清洁 ³ ，王子菊 ¹
AuthorsListE:Shi Zaifeng,Li Huilong,Cheng Ming,Cao Qingjie,Wang Ziju
AuthorsHTMLE:Shi Zaifeng^{1, 2}，Li Huilong¹，Cheng Ming¹，Cao Qingjie³，Wang Ziju¹
Unit: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 300387

Unit_EngLish: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

Abstract_Chinese:低剂量双能计算机断层扫描成像(DECT)技术可以在提供人体内部结构及组织成分信息的同时减少 X 射线 辐射剂量. 然而，剂量的降低会导致 DECT 重建图像中出现大量的噪声及伪影，从而影响对疾病的精确诊断. 为实 现在低剂量条件下重建出高质量的 DECT 图像，提出了一种采用混合高斯模型的正弦图去噪网络来进行伪影及噪声 消除. 该网络由两部分构成：一部分通过残差学习以有监督的方式对校准后低剂量与正常剂量下 DECT 投影数据的 映射关系进行拟合；另一部分采用混合高斯模型以无监督学习的方式提取 DECT 投影数据中噪声的分布模型. 采用 这种监督与无监督学习结合的方式，不仅可以利用卷积运算的特征提取能力来拟合输入与标签之间任意复杂的映射 关系，还可以在无标签约束的情况下，利用输入投影数据的自身分布规律来提高网络模型去噪性能及其泛化能力. 实验使用了 XCAT 生成的 10 名不同人体 DECT 投影数据对网络模型进行训练及测试. 实验结果表明，与正常剂量 下获得的重建图像相比，该方法所获得的去噪后图像均方根误差值低于 6×10-3 ，峰值信噪比以及结构相似性指数分 别超过 36.7 dB 和 0.992. 相比于目前先进的低剂量 CT 噪声去除方法，该方法得到的 DECT 重建图像中组织结构更 加清晰，并且可保留更多的细节信息，可为后续医疗诊断提供精准参考.
Abstract_English:Low-dose dual-energy computed tomography(DECT)has the potential to provide information on human internal structure and tissue components and to reduce X-ray radiation. However，dose reduction often leads to extreme noise and artifacts in reconstructed images，which dramatically affects the accuracy of the diagnosis. In order to obtain high-quality reconstructed images from low-dose DECT projection data，a noise reduction network called sinogram denoising network with Gaussian mixture model(SDN-GMM)was proposed to eliminate artifacts and noise. Further，this network consists of two learning parts：the supervised and the unsupervised. In the supervised learning part，the relationship between calibrated low-dose and normal-dose projection data was determined by residual learn\u0002ing，while the unsupervised learning part extracted the noise distribution of DECT projection data via Gaussian Mix\u0002ture Model. The combination of supervised and unsupervised learning not only can take full advantages of the featureextraction capability from convolution operation to suit any complex mapping relationship between the input and the label but can also make full use of the input data property to further enhance the efficiency and robustness of the net\u0002work model. In the experiment，the DECT projection data from 10 different people acquired from XCAT were used to train and test the proposed network model. Compared with the normal-dose reconstructed images，the results revealed that the root-mean-square error(RMSE)value is lower than 6×10-3，and the peak signal-to-noise ratio(PSNR)and the structural similarity index measure(SSIM)are higher than 36.7 dB and 0.992，respectively. On the other hand， compared to the current advanced low-dose CT noise reduction methods，the DECT reconstructed images produced by proposed method have clearer tissue structure and can retain more detailed information，which will be more valuable for medical diagnosis.
Keyword_Chinese:双能计算机断层扫描成像；低剂量；残差学习；无监督学习；混合高斯模型
Keywords_English:dual-energy computed tomography；low dose；residual learning；unsupervised learning；Gaussian mixture model

PDF全文下载地址:http://xbzrb.tju.edu.cn/#/digest?ArticleID=6675