心脏应力场建模是心脏影像分析的核心技术，能为心脏疾病的定量解释、诊断与预防提供准确全面的依据。无标记四维心脏CTA和电影MRI蕴含着细微的三维动态应力场特征，但受影像模糊、组织密度混叠、目标运动复杂、机器学习技术挖掘目标先验信息能力不足等因素影响，优质建模方法极为匮乏。对此，该项目研究可解释性强的深度流形度量学习(DMML)新方法并应用于心脏应力场建模关键环节，通过融合DMML与心脏形态运动先验的深度表达来提高建模精度。内容包括DMML算法、四维心脏图谱集构建、心脏粗分割、多目标迭代软分割、运动估计，着重对适用于心脏图像的流形学习和图嵌入、嵌入流形学习卷积层的深度孪生网络、心脏分割与运动估计中边界图像相似性度量等关键技术展开系统深入的研究，注重利用多相CTA/MRI图像验证和改进算法。该课题对延拓深度度量学习理论内涵与新型应用，推动心脏建模发展有积极作用，也能为心脏病诊断提供技术支持。

It is a key and challenging technique of building the stress filed of left ventricular and its myo-/endo-/epi-cardium muscles in the research area of computer-aided cardiac imaging analysis and diagnosis, since this field can express the underlying, objective and quantitative diagnostic indices and parameters to help the explanation of morphological variations, diagnosis, and tracking of the cardiac ventricle change. Due to the facts such as rather obscure beating heart, highly variable structures, and mixed organs during the image capturing and reconstruction process, as well as the absence of advanced machine learning-based strategy to thoroughly explore the geometric and movement prior information in most traditional methods, existing stress field models are usually of low accuracy. To address this, in this proposal, new deep manifold metric learning methods (DMML) with good interpretable characteristics will be investigated and applied to cardiac ventricle stress field modeling of the four dimensional untagged CTA (computer tomography angiography）and cine MRI (magnetic resonance imaging) data sets. The characteristic of this proposal lies on combining DMML and prior information (such as the geometry, shape, and motion) in cardiac image data sets to advance the development of accurate ventricle stress field modeling. This research proposal mainly includes the exploration of DMML training and modification specific to the cardiac anatomic and imagery structure for improving the accuracy in cardiac stress modeling, four dimensional cardiac atlas set construction, cardiac multiple objects coarse segmentation their iterated soft segmentation, and motion estimation. The shape or motion prior information will be built by the libraries of shape image or motion image trained by deep metric learning strategy. The investigation will focus on the technologies related to efficient manifold learning and graph embedding techniques specific to cardiac images, the deep metric networks such as siamense networks that contain the manifold learning-based convolutional layers, the coarse and iterated soft segmentation of multiple cardiac tissues, as well as the endocardium and epicardium motion estimation in this region because they essentially determine the quality of the stress field. It will also focus on using the practical CTA and MRI data with multiple time-points to validate and improve the technique and corresponding algorithms. Both theoretical and practical achievements will be obtained if this proposal is supported. Theoretically, new deep manifold metric learning methods will be presented, and the solution for cardiac segmentation, motion estimation, stress field analysis in cardiac modeling methodology will be put forward with large steps. Practically, the research algorithms and software can provide assistance in daily cardiac diagnosis clinic, cardiac surgery simulation, and statistical analysis on cardiac disease.