三维MR影像中非均质柔性生物组织内部变形场的高精度测量方法研究

Current registration methods of medical images cannot steer the accurate matching of a number of feature points in the deformation measurement of nonuniform and nonrigid biological tissues. Based on the study of the extraction and matching of the feature points, we would research the accurate deformation measurement of the nonuniform and nonrigid biological tissues from reconstructed 3D MR images. Firstly, we would research the method of feature point detection from the deformed 3D MR images,such as the Laplacian of the Gaussian (LoG) regarded as the robust feature point detector. Afterward, we would enhance the detection accuracy and the scaling robustness of the Harris by using the B-Spline approximated filter and the pyramid image, respectively. Then the LoG and improved Harris are used to detect a number of the blob feature points and corners spreading all over the biological tissues symmetrically in the MR images. Secondly, we would research the descriptor of feature point which is robust to the nonlinear deformation of the 3D MR images. According to our proposal, the image patch of 3D feature point is mapped to the 4D Riemannian manifold. Then we would describe the Riemannian manifold of feature point by heat diffusion geometry, and the feature point descriptor is built by the heat kernel of the heat diffusion equation. Afterward, the feature points are matched by this descriptor being robust to the nonlinear deformation. Thirdly, mismatching of the feature points would be eliminated automatically by the method combining the 3D TPS model and the preserving of the local topology around the feature point. Then we can implement the accurate matching of a number of feature points between the initial and deformed 3D MR images. Finally, TIN (Triangulated Irregular Network) would be built based on the matching feature points. The displacements of the feature points and the interpolated points of the TIN are calculated accurately by tiny facet primitive rectifying. By the abovementioned methods, we would implement the accurate deformation measurement of nonuniform and nonrigid biological tissues from the 3D MR images. We would solve some key problems, such as the accurate matching of numbers of feature points, eliminating the mismatching automatically, in the deformation measurement of nonuniform and nonrigid biological tissues which would be widely used in the modeling and simulation of the biological tissues, design of the nonrigid robot, and so on.

针对当前医学影像配准等方法难以解决非均质柔性生物组织内部变形场测量中大量特征点准确匹配的问题；项目基于非均质柔性体非线性形变前后的三维重构MR影像，拟采用稳定性较好的高斯Laplacian与B样条函数滤波Harris算子，提取MR影像中大量均匀分布的特征点。并将三维特征点映射到四维黎曼空间，采用热扩散方程对特征点黎曼流形进行描述，构建非线性形变鲁棒性特征点描述符，对特征点进行描述与匹配。同时，针对可能存在的误匹配点，利用三维TPS模型并结合局部拓扑结构保持的策略进行剔除，便可实现MR影像中大量特征点的精确匹配。在此基础上，根据匹配的特征点构建TIN三角网，用以计算柔性体形变时特征点的位移，以及TIN内部插值点的位移，实现柔性体内部变形场的高精度测量。本项目研究成果，将使非线性形变MR影像中大量特征点的精确匹配等关键问题的解决有较大的突破，并推进其在生物组织建模与仿真等领域的广泛应用。

针对当前医学影像配准等方法难以解决非均质柔性生物组织内部变形场测量中大量特征点准确匹配的问题；项目研究了非均质柔性生物组织三维MR影像中特征点的检测算法，包括Harris、DoG、Fast Hessian特征点检测算法等。研究了特征点结合B样条的FFD(free-form-deformation)模型层间影像配准方法，实现MR切片间影像的匹配与插值。特别是针对非线性形变MR影像中大量特征点匹配的关键问题，研究了构建非线性形变鲁棒性特征点描述符、特征点的精确匹配，误匹配点的自动剔除算法等。提出了TSSC（TPS-SURF-SAC-Clustering）非线性形变特征点匹配与误匹配自动剔除算法，针对TSSC初始条件下需要选择少量标志点的问题，进一步提出了将SIFT特征点描述符与局部邻域结构保持策略相结合的特征点描述与匹配方法SPLNS（SIFT and Preserving Local Neighborhood Structures）。为更精确描述特征点的纹理、空间拓扑结构信息等，提出了SPTSLN (SIFT and Preserving Topology Structures of Local Neighborhood)算法。上述算法相对于经典的SIFT、SURF等匹配算法而言，具有更高的精度，并能够获得更多的正确匹配点对。鉴于黎曼流形对变形描述方面的优势，研究了基于黎曼流形非线性形变特征点匹配算法。将特征点通过拉普拉斯特征映射流形学习方法映射到黎曼空间，采用热扩散方程对特征点黎曼流形进行描述，构建特征点描述符，对特征点进行描述与匹配，但是该研究结果并不理想。最后研究了基于匹配特征点的Denaulay非规则三角网构建方法，根据小面元纠正的位移场测量方法。研究成果形成了相应的论文，并申请了一项软件著作权。后期将继续针对黎曼流形非线性特征点匹配算法进行深入研究。