基于联合显著图和局部自适应核回归的非刚性医学图像配准及临床应用研究

To meet the urgent demands for tracking and quantifying the lession tissue changes in modern clinical application, nonrigid registration of multimodal and multitemporal images often meets great challenges due to the great geometric difference between the images introduced by the local large deformation and missing correspondence (or outliers) of local tissues in the same scene. In order to tackle these problems, this project proposed following novel methods: .1) By adaptively learning the local structures and their similarity measures from the two images to be registered, we propose a novel joint saliency map (JSM) as a new structural similarity measure in nonrigid registration to distinct the non-correspondence structures from correspondence structures and backgournd, with the JSM representing registration bias in these non-correspondence structures to iteratively reduce registration errors and boost the registration performance in registration feedback iteration..2) We further introduce a novel nonrigid registration framework, based on joint-saliency structure adaptive kernel regression, to aim at handling local complex large deformation and outliers. The sparse displacement via multi-resolution block matching is locally and adaptively smoothed according to the structural adaptive kernel function which is adaptively learned from local anisotropic structures and their local similarity measures. Moreover, the joint saliency map reflecting structural correspondence and registration bias across the two images guides the assignment of data certainty to the reliable corresponding structural voxel pairs that are emphasized in the local kernel regression of deformation fields. .3) To meet the recent research interest in most successful non-parametric approaches in image processing and fit for the purpose of accurately matching small local structures with smooth deformation in some clinical applications with tumor change and cardiac perfusion images, this project proposes several research ideas, i.e., low-dimensional manifold kernel regression and multivariate kernel regression with different choice of basis function and kernel function, to be deeply involved in the locally adaptive kernel regression (nonparametric regression) combined joint saliency map for local smooth regularization of intensity-based nonrigid registration, whereby our proposed methods surpass traditional intensity-based and feature-based methods that are limited in accurately matching local small structures.

多模式多时段医学影像技术的发展，和临床对肿瘤病灶、心肌灌注成像等复杂图像定量分析的紧迫需求，促使本项目要解决好在局部大形变和组织对应性缺失的复杂异常条件下、现有非刚性医学图像配准算法大都失效的问题：1)基于待配准图像局部显著结构及其相似性表示构建联合显著图，准确区分已配准好结构区域、配准残差大的局部大形变和对应性缺失结构区域以及背景区域。进一步提出新的结构相似性配准测度和配准残差减少反馈机制，来迭代提升非刚性医学图像配准精度。2)利用多尺度子块匹配对局部各向异性尺度、方位和局部相似性测度的自适应学习，以及各像素对基于联合显著图的"确定度"分配，局部自适应核回归无监督地精确配准医学图像的局部显著结构(或连贯分布图像特征)。3)从自适应核回归基函数和非对称核函数、多点逼近多元核回归、以及低维流形核回归等角度，综合研究非刚性医学图像配准的非参数化处理新理论和新方法，顺应非参数化图像处理研究趋势。

本项目对同时具有局部结构大形变和对应性缺失的复杂异常图像、现有非刚性图像配准算法大都失效的问题，结合临床病灶发展、心脏冠脉灌注造影成像等复杂图像定量分析的需求，展开和临床密切合作的基础应用研究。1）通过匹配基于结构张量的边缘感知显著图分布，构建出待配准图像的边缘感知联合显著图，突出显示待配准图像重叠区域内对应性显著结构(称为联合显著结构)的边缘匹配情况。由此，我们开创了包括基于固定尺度、局部尺度以及结合深度学习在内的联合显著结构自适应核回归(JAKR)方法体系进行非刚性图像配准。2）JAKR认为图像配准是图像每个体素形变场的重建过程，在对每个体素通过其邻域稀疏形变场核回归重建为致密形变场的过程中，一方面调整各向异性核形状和方位能自适应参考图像的结构，另一方面突出联合显著结构的形变场在核回归中的贡献权重。这样，浮动图像的局部形变最大程度地匹配参考图像的每个对应结构，同时又有效抑制了局部大形变和对应性缺失的不利影响。3）相比于固定尺度核回归，局部尺度JAKR (LSJAKR)采用的各向异性核窗口大小不再是固定的，而是根据图像内待匹配局部结构尺度以及基于边缘感知联合显著图的边缘不匹配程度进行自适应调节的。相比于当今其他先进算法，JAKR对复杂异常图像的配准精度最高，尤其是LSJAKR更能对微小结构做到最为精准的配准。本项目又结合深度学习发展的深度JAKR方法，更能高效提升非刚性图像配准的计算性能。4）基于心脏冠脉灌注造影成像背景结构的相似性低秩建模和图像分解，本项目提出了运动一致性约束累进鲁棒主分量分析法从复杂的造影图像背景中提取出X射线造影血管，为各种基于造影血管的临床诊疗技术奠定了可靠和必要的技术基础。5）项目还拓展提出了自助法泊松回归图像降噪和背景/前景分类约束的非负矩阵分解谱解混等方法，有效提高荧光谱成像的定量分析精度和临床应用水平。