我想认识我自己
四室心脏模型和自动分割
Four-Chamber Heart Modeling and Automatic Segmentation for 3D Cardiac CT
Volumes Using Marginal Space Learning and Steerable Features
Authors: Yefeng Zheng, Adrian Barbu, Bogdan Georgescu, Michael Scheuering, and Dorin Comaniciu
There are two topics to discusse: heart modeling and automatic model fitting to an unseen volume. The first topic propose an approach based on control points, which are integral part of the mesh model in the sense that they are also connected to other mesh points with mesh triangles. In the heart model, the mitral valve is explicitly represented as a closed contour along its border and they exclude the moving valve leaflets from the model, therefore the basal area can be delineated more consistently. Both the endo- and epi-cardiums are delineated for the LV. Since the RV has a complicated shape with separate inflow and outflow portions, they split the RV into three parts with the RV divergence plane, which passing the divergence point of the inflow and outflow tracts(the cyan line in the picture). In the RV the two cusps are important landmarks. In the establishing point correspondence, the paper uses different methods to select cutting plane, they develop two resampling schemes, the rotation-axis based method for simple parabola-like shapes and parallel-slice based method for the more complicated RV. In both methods, the long axis of a chamber is defined the long axis as the line connectiong the center of a valve and the mesh point farthest from the valve. The rotation-axis cuts the mesh with a plane passing the axis and the parallel-slice uses a plane perpendicular to the RV long axis to cut the 3D mesh.
(The picture is coming from the paper)
The second topic is about 3D object localization or detection. They use the marginal space learning (MSL) and steerable features to improve the efficiency. MSL formulates the localization problem as a classifacation problem. We train the classifier by a set of CT volumes, for each volume a nine dimensional vector of the ground truth, which are position, orientation, and scaling of the heart chamber, are needed.The classifier can assign a score for each hypothesis, they select the best one or several hypotheses. The MSL has three steps: position estimation, position-orientation estimation, and full similarity transformation estimation, for avoid the exponential increase. To train the classifier, the hypotheses' ground truth are splited into positive and negative by a normalized distance measure (which defined by normalizing the error in each dimension to corresponding search step size ). Training of position estimator's search step is one voxel and use 3D Haar wavelet features for learning, training a classifier using the probabilistic boosting-tree (PBT) in this part. In the following, training of position-orientation estimator and similarity transformation estimation learning and training in the same way, the search step is 0.2 radias(11 degrees), 2 voxels(6 mm),respectively. The dimension of candidates is also augmented.
For the steerable features, the orientation and scale information are enbeded into the distribution of sampling points, while each individual feature is locally defined. The sampling pattern is steered rather than aligned the volume to the hypothesized orientation.
At last the mean shape is deformed to fit the object boundary. There is 3 steps: first, estimating the deformation of control points. Then warpping the initial mesh toward the refined control points by the thin-plate-spline(TPS) model. Last, the normal mesh points are deformed to fit the image boundary.
我写了一篇自己都不会看的烂读后感。。。。还是再写一遍中文的吧。。。。
这篇文章讲了个啥事呢，其实就是讲了他们建立了一个四室心脏模型，然后用这个模型做了一个基于3D CT数据的自动分割系统。
这个四室心脏模型是咋做的呢，其实就是通过标定各部分之间的控制点，将点连成线形成分割面。这个点是怎么设定的呢，就是通过在曲面上进行重采样的方法，对于对称的结构（如左心室，左心房，右心房）采用旋转轴进行分割；对于不对称的结构（如右心室）采用平行切面的方式进行分割，分割面垂直于长轴。对分割出来的切面轮廓进行采点，即控制点。长轴的选取定为连接瓣膜中心点和距离瓣膜最远的网格上的点的连线（描述了个啥。。。）。在右心室右心房建模中吗，把右心室分割为三个部分的主区域、右心室流出道和右心室流入道。右心室主区域的分割面是新月形的，新月两头的尖角点是重要的探测对象，需要被探测出来。
3D物体定位被视为一个分类问题，使用的是MSL和可控标志。分类器的训练数据是CT数据集，每个数据集的真值是一个九维向量，坐标T（x,y,x）、方向（ψ,φ,θ）（方向表示为一个旋转矩阵R）和伸缩比例（Sx,Sy,Sz）。坐标系：心室的长轴记为x轴，垂直于长轴记为y轴，原点为长轴上预先设定的锚点（不同的心室锚点选择不同），z轴是垂直于xy平面的，参考系选择是欧拉角，zxz顺规。各结构的方向已知，只需将它从世界坐标系转换为物体坐标系中，然后在整个训练集中计算其均值图形.转换公式如下。 
MSL：设计一个分类器逐步筛选最佳假设，输入为所有假设，分类器为每个假设打分，输出为分数高的一个或者几个假设（分数高低个人设定）。分类器分三个阶段对假设进行筛选，依次为坐标、坐标和方向、全部的相似变换。每一阶段都选择一些候选假设以降低搜索空间。
分类器的训练：基于他们的真值将所有假设分为积极和消极两组，根据错误率进行分类，错误率计算机分类方法如下。
三个向量的训练方法基本一致，使用3D哈尔小波特征进行学习，Probabilistic Boosting-Tree（概率题高树）训练分类器，然后对假设进行分类。
1）位置训练使用的步长是1体像素，
2） 位置-方向步长是0.2弧度（11度），扩展数据集的维度至六维（位置+方向）
3）全部相似变换训练步长是2体像素（6mm），数据集维度均匀而详尽的扩展至9维。
Steerable Features
全局特征（如3D哈尔小波特征）可以很有效的获取全局信息，但是为了获取位置信息我们需要旋转特征模板，然而并没有有效的方法去旋转特征模板，局部信息又会丢失全局信息。进而需要导向特征，导向特征是同时获得方向和伸缩比例信息。
对于每一个采样点从原始数据中基于强度I和梯度g提取一些局部特征。
对心脏定位后需要使均方形状变形使其与真实边界进行匹配，匹配过程中使用的是导向特征。匹配过程分为三步。
1） 估计控制点的变形。（使用MSL探测，获得控制点的均方形状）
2） 使用薄板样条（thin-plate-spline TPS）删减初始网格以使控制点更好的校准。（使用边界探测器沿着法线方向移动探测每一个控制点）
3） 变形网格去匹配图像边界，控制点不改变。（把变形的图像映射到一个形状子空间，形状子空间个人设定，这篇文章设定的是可以包括98%的变异）
文章的实验过程：
数据集缺陷：不能研究相同或不同病人的变异。
检测方法：四倍交叉检验。
在边界勾画时独立测试每个结构。这个系统测试了的数据集包括了心脏的所有周期。
这是我读的第二篇文章，第一篇和这个是一系列的，第一篇介绍的是这个模型的应用。初次踏入这个领域，谈不上对于这个方法的评价，因为没有对比呀。只是觉得好复杂。。。。。。这个控制点到底是咋提取的，咋获得的图像边界呢（哦 好像是每一个数据都有专家备注的，这个样子呀）好多不认识的知识，加油啦！