医学图像范例先验构造与虚拟多模态成像方法研究

Virtual multi-Modality Imaging (VMI) can synthesize the images of a target modality only using the image data of a source modality by simulating the real imaging devices, and can provide the potential and important visual information of the target modality. However, the existing image prior models can not be effectively applied to solve the highly ill-posed modality synthesis problems of VMI due to the difficulties in learning their model parameters or their limited constraint ability. To obtain sufficient prior knowledge for developing the high-quality VMI technology, this project aims to establish the methods for building the exemplars-based prior models of medical images. Specifically, the non-parametric exemplars-based image prior models for the specific modalities and anatomic sites are built by utilizing image retrieval and image patch matching techniques on the basis of large-scale medical image database. The research priorities of this project include: 1. design of the search and selection strategies of the exemplar images, realizing effective sampling of the image neighbors; 2. design of the hierarchical matching algorithms to obtain the correspondence fields between the medical images, realizing efficient sampling of the image patch neighbors; 3. the modality synthesis models with the exemplars-based priors for virtual CT imaging from MRI image data (vCT-MRI) and virtual Dual Energy Subtraction (vDES). It is to be hoped that the high-quality vCT-MRI and vDES can be achieved through the proposed methods, and can be applied to PET attenuation correction in PET/MRI system and separation of the overlapping anatomical structures in X-ray chest radiographs, respectively.

虚拟多模态成像（Virtual multi-Modality Imaging, VMI）是指模拟医学成像设备功能，由源模态图像数据合成特定模态图像的技术，实现医学图像中潜在重要影像信息的提取。高质量的VMI模态合成需获取有效的先验知识，而现有图像先验模型尚不能实现这一目的。鉴于此，本项目提出基于范例图像的医学图像先验（范例先验）模型构造方法，以大规模医学影像数据库为基础建立针对不同模态、不同部位医学图像的非参数先验模型，研究重点包括：（1）设计范例图像搜索和优选策略，实现图像近邻的有效采样；（2）设计层次匹配方法高效求解图像对应场，实现图像块近邻的高效采样；（3）构建结合范例先验的VMI模态合成模型，以期实现优质虚拟CT-MRI成像和虚拟双能减影，用于解决PET/MRI系统中PET图像衰减校正和X线胸片中重叠解剖结构影像分离的问题。

本项目针对虚拟多模态成像问题中先验知识的有效获取方法进行研究，以期实现高质量的vCT-MRI（由MR图像数据预测合成CT图像）和vDES（胸片虚拟双能减影），用于解决PET/MRI系统中PET图像衰减校正和X线胸片中重叠解剖结构影像分离的问题，具有重要应用价值。本项目研究按照任务计划书实施，提出了图像非线性局部描述子学习优化方法，用于范例图像搜索和图像块匹配；提出了图像对应场求解的层次匹配求解方法和图像块匹配的局部稀疏匹配方法，用于构造范例先验；提出了多层次锚点近邻回归模型实现vCT-MRI，并通过图像数据驱动优化方法提升预测CT图像在空间上的一致性；提出了迭代加权最小二乘的优化算法优化求解基于范例先验的胸片图像分解模型，有效实现了胸片vDES。. 本项目共发表相关学术论文10篇（SCI收录7篇），其中2篇分别发表于医学影像处理分析领域顶级期刊IEEE Transactions on Medical Imaging和Medical Image Analysis、1篇发表于核医学领域顶级期刊Journal of Nuclear Medicine（影响因子：6.65）、1篇发表于生物医学工程领域主流期刊IEEE Journal of Biomedical and Health Informatics；申请国家发明专利5件、其中3件已授权；开发完成deepVDES胸片骨抑制软件，并在南方医院放射科进行试用和评估；培养硕士生4人、博士生4人，较好完成了任务计划书中的计划目标。