面向低剂量能谱CT的对抗学习式能量探测与物质分解技术研究

能谱型X射线计算机断层扫描(CT)技术可通过X射线中各能量成分的衰减情况进行断层图像重建，获得更精准的影像信息，已广泛用于现代医学临床诊断。X射线辐射会对人体健康造成不利影响，研究如何在降低辐射剂量的同时保证重建图像质量是该领域的热点和难点。本课题拟研究低辐射剂量下能谱CT在探测成像和物质分解中的串扰、数据缺失和噪声等问题。课题创新性在于：（1）通过建立级联分析模型结合对抗式学习方法，以校正低剂量下引起的能量探测串扰误差；（2）利用双生成-鉴别循环的无监督循环一致对抗学习网络，实现CT投影数据的补全；（3）利用与多能量区间相对应的多生成器的GAN网络和复合损失函数进行物质分解过程中的噪声消除。课题拟突破当前低剂量下能谱CT的技术瓶颈，利用生成对抗式深度学习技术，从X射线能量探测串扰校正、投影数据补全和物质分解噪声消除三方面进行技术攻关，最终形成一套低剂量能谱CT能量探测与物质分解方法学。

X-ray computed tomography (CT), especially spectral CT, has been widely used in modern clinical diagnosis, and it can obtain accurate material information by measuring the linear attenuation coefficient of scanned objects under different energy bins. However, the high radiation dose of X-ray is harmful to human health. To reduce the X-ray radiation dose and maintain clinically satisfied CT image quality at the same time is the key issue in this filed. In this study, we focus on solving the problem caused by dose reduction, including the incomplete data, amplified noise and interlayer crosstalk. The novelty of this study is as follows. (1) By establishing a cascade analysis model combined with an adversarial learning method, the energy detection crosstalk error caused by dose reduction is corrected. (2) The missing CT projection data are restored through employing an unsupervised circular uniform adversarial learning network with dual generation-discrimination cycles. (3) A generative adversarial network with multiple generators and loss functions corresponding to multiple energy bins is used to eliminate noise during the material decomposition. This study intends to break through the current technical bottleneck of the spectral CT, and use the generative adversarial deep learning method to carry out technical research from three aspects, including X-ray energy detection crosstalk correction, projection data completion and material decomposition noise reduction. Finally, a novel methodology of energy detection and material decomposition for low-dose spectral CT will be developed.