基于深度学习实现能谱CT成像的物理模型和大数据结合优化

Spectral CT is a state-of-art technology in X-ray imaging. It breaks through the physical limits of traditional single energy CT with unprecedented material identification and K edge imaging capabilities. It is suitable for low-dose imaging, quantitative imaging, as well as dangerous goods testing for security inspection. Before its comprehensive application in the medical, industrial and public security fields, it faces fundamental constraints from the energy-selective photon-counting detectors/imaging physics and computational complexity, which could cause the failure of cashing the advantages of spectral CT. These aspects greatly affect the promotion of spectral CT imaging systems. This project will research on an innovative design that combines this latest imaging model and deep learning, the most cutting-edge technology in computer vision to optimize the imaging performance of spectral CT: By adapting the struction of convolution neural network as the basis framework, we are to build high-performance deep neural networks for an innovative spectral CT data processing system, including networks for detector modeling and energy spectrum calibration, multi-dimensional spectral CT reconstruction, spectral CT denoising/artifact suppression. We will introduce physical layers, statistical models, and regularization constraints, in the neural network design to achieve a combined optimization of physical model and big-data-driven so that integration of cross-field advantages can be gained. In this research, the current isolated processing mode with single sets of data of spectral CT will be exceeded. From deep mining the large volume data for build-in features and patterns, we would be able to break through the information limit in a single set of data. Moreover, the new methods would not only improve system performance in image quality, but also transfer large amounts of computing to network training in preprocessing so that greatly enhance the actual processing speed after system deployment. Overall, this project is a pioneering exploration of a new research direction in the X-ray imaging field. Its achievement will greatly enrich CT imaging theories and technologies.

能谱CT突破了传统CT的物理极限，具有前所未有的物质成分识别和K边缘成像能力，适用于低剂量成像、定量成像、安检危险品检测等应用。在其全面走入医学与工业应用前，还面临来自探测器和成像物理的根本性制约和高计算复杂度造成的优势兑现困难，影响系统的推广。本项目将研究最新成像模式与计算机视觉最前沿技术深度学习结合优化的创新设计：以卷积神经网络为主架构设计新型能谱CT数据处理系统，包括建立探测器建模和能谱标定、多维度能谱CT重建、去噪/伪影抑制的高性能深度网络，通过在网络设计引入物理层、统计模型、正则化约束，达到物理模型和数据驱动的结合，实现交叉领域的优势综合；打破当前能谱CT单套数据的孤立处理模式，从大数据集深度挖据特征，突破单套数据的信息量极限；不仅实现系统性能提升，而且把大计算量转移至网络训练，大幅提升系统部署后的实际处理速度。项目的研究将开启X光成像领域新的研究方向，丰富CT成像的理论和技术。

X射线能谱CT依据能量分辨信息提取包括有效原子序数、电子密度、单能线衰减系数等多维信息，在医疗诊断、安检危险品检测等具有广阔应用前景，已经成为X射线成像领域近年来的最大变革性发展。本项目完成了此最新成像模式与计算机视觉最前沿技术深度学习结合的数据处理最优化方法研究，在功能模块上包括能谱CT数据库建设、能谱CT系统标定和能谱信息分解设计、伪影抑制网络的设计、端对端多功能网络的设计集成和优化、能谱CT系统的优化方法，总体建立了系统的基于数据驱动的深度学习能谱CT成像的图像重建和系统性能优化方法。项目首先从X射线能谱CT成像过程的细化物理模型出发，联合精确算子的网络层嵌入和跨域误差反向传递，建立了单域学习和多域综合学习的端对端网络深度学习的理论方法和统一设计框架；其次，从全新CT专用代价函数设计、能谱CT信息分解理论指导下的网络架构精细化适配突破了常规神经网络的性能和效果上限；进而，通过创新基于模型和深度分布约束的无监督学习、CT特有的数据域迁移和对齐方法、三维信息解构，打通了实际应用壁垒；最终形成了完整地适用于能谱CT成像和普通CT成像的深度神经网络设计和应用方法，以及先进高效的实施技术。本项目深度实现了物理模型和数据驱动的结合、交叉领域的优势综合，达到对比传统方法的全面性能提升，相较领域内近年提出的其它深度学习类方法在方法的可解释性、鲁棒性、实际应用性能亦具有优势，圆满地完成研究目标。研究成果开拓了贯通传统理论积累和前沿新方法的新路线，丰富了CT成像的理论和技术，也为X光成像领域先进设备研发提供了坚实的方法和技术支撑，兑现显著应用价值指日可期。