面向三维图像配准与属性迁移的无监督聚类森林测度学习方法研究

3D image registration and attribute transfer via the unsupervised clustering forest with metric learning plays an important role in 3D image superimposition, operation planning, treatment evaluation, and the assessment of 3D structural growth. In our research project, we will study the following major aspects: the generalized metric learning algorithm based on the unsupervised clustering forest, the supervoxel-wise consistent correspondence between 3D volume images, the theoretical complexity analysis of the unsupervised clustering forest, and the end-to-end 3D structure parsing. We will present a clustering forest-based metric, where the weak labeling is utilized to discriminate the badly clustered supervoxels and related classes. The clustering forest is enforced by penalized weighting in both classification and clustering entropy estimation using the weak labeling. We will discuss consistencies of candidate matching pairs by isometric constraints, which screen out the inconsistent matching pairs by regularizing the dense correspondence defined by the forest-based metric. We will present a theoretical complexity analysis with a data-dependent learning guarantee for the clustering hypotheses of the unsupervised clustering forest considering both the sequential tree traversals determined by questions stored in branch nodes and the clustering compactness in leaf nodes. A novel tree-pruning algorithm is proposed to refine the forest by removing local inconsistent leaf nodes. We will study a fully automatic structural parsing of the CBCT images by a volumetric convolutional network with nested residual connections using multi-scale feature fusion and multi-level deep supervision, as well as a segmentation feedback scheme. The project will focus on two scientific challenges to provide the theoretical bases and the core algorithms for the relevant applications. The first one is learning mechanism of the clustering forest-based metric. The second one is the online computationally efficient framework for automatic registration, attribute transfer, and structural parsing of 3D volume images.

面向三维图像配准与属性迁移的无监督聚类森林测度学习方法研究，将为临床治疗中三维图像重叠、手术预测、疗效评价、三维结构生长评估等提供重要的技术支撑。研究基于聚类森林的测度学习方法，设计基于弱标注的分类以及聚类信息增益融合技术，迭代优化聚类森林并建立面向三维图像的测度；研究基于聚类森林的一致对应方法，提出柔性一致对应评价机制，对稠密对应进行筛选和重建，以避免传统方法中基于非线性优化的正则化等耗时计算；研究无监督聚类森林模型理论复杂度，利用Rademacher复杂度获取聚类森林数据相关假设函数族上界，为聚类森林的构建提供理论支持；研究端到端的三维图像结构解析技术，结合多尺度特征融合与多级深度监督技术，利用回馈机制，以提高现有三维图像结构解析性能。将重点探索无监督聚类森林测度学习机理，以及面向三维图像属性迁移与结构解析的框架研究，为临床三维医学图像处理提供核心算法与理论基础。

面向三维图像配准与属性迁移的无监督测度学习方法研究，对临床三维图像重叠、手术预测、疗效评价等提供重要的技术支撑，蕴含重要的理论与应用价值。项目研究内容包括（1）面向高维数据相似度估计的无监督聚类随机森林测度，设计基于聚类森林的连续度量与伪叶分裂方案，缓解不一致的叶节点分配与对应，实现基于聚类森林测度的三维图像配准；（2）面向配准与体素属性迁移的三维图像特征表示学习方法，结合低级视觉局部特征表示与高级全局关联，设计三维颌面图像嵌入算法，探索跨维度双向离散嵌入映射，实现三维图像配准与结构对应；（3）三维图像配准中的一致结构对应方法，设计双向可逆与闭环一致约束，探索迭代反馈机制，构建双路深度自注意力网络架构，基于全局和跨图像相关建模改进解剖结构的语义一致对应；（4）研究端到端三维图像结构解析方法，设计实例感知的测地距图用于上下文空间关系编码，构建无监督跨域结构保持图像生成框架，实现有限标注下的三维颌面结构自动标注。项目在相关领域都取得了成果。在算法理论研究方面，团队提出了基于聚类森林的测度学习方法，设计了面向配准与体素属性迁移的三维图像特征表示学习与一致对应方法，构造了基于有限标注样本的三维颌面结构自动标注算法，改进了三维图像配准与结构解析精度，在相关领域发表了高水平论文，促进了相关学术领域的发展。在应用方面，研究团队在针对临床正畸病人由于治疗以及发育造成的结构变化的定量评价、牙合结构度量、特定病人三维结构预测等应用中取得了成果并申请专利，服务于智能临床三维信息辅助诊疗。