体会MIMICS图像分割

jati

我看到这个论坛一直很少讨论分割，就把毕业论文里一部分讲述图像分割的文字，按照MIMICS的功能点，改写并补充成一篇教程，是我的体会，不是官方说法。凑合看吧～




图像序列一旦导入到MIMICS中，就将其视作体数据场。MIMICS从体数据场中获得一个mask（目标数据点集、面罩），并将这个mask用于生成三维模型。根据体数据场获得mask的过程就是图像分割。图像分割是可视化、有限元建模等后续工作的基础，是从断层图像数据展开医学有限元、医学快速成形、医学虚拟现实等应用的关键步骤。MIMICS的Segmentation菜单下提供了一些图像分割工具，能够实现最基本的分割算法（除了微分算子）。MIMICS的图像分割是交互式进行的，因此分割效果很大程度上取决于手工分割的流程设计。下面介绍Segmentation菜单下的几种重要的分割工具：
Thresholding
提取指定灰度范围的数据点集构成一个mask，是最常用的分割工具。由于这个方法参照体数据场本身的信息，所以往往是分割的第一步。
RegionGrowing
根据已知的mask和指定的种子点，得到种子点所在的相邻数据点集作为mask。这个方法并不参照图像本身的信息，只参照目标mask和种子点。也是很常用的分割工具。
DynamicRegion Growing 根据指定的种子点，得到在指定阈值范围内的数据点作为mask。既参照了阈值信息，也参照了指定的种子点。MIMICS的广告彩页称这个工具是分割神经、血管等组织的有效工具。
MorphologyOperations
对mask进行腐蚀、膨胀、开运算和闭运算得到新的mask。形态运算并不以图像本身信息做参照，仅仅是针对已知mask进行的运算。形态学工具对于一些噪声很有用，这个算法具体做了什么，可参阅《医学图像处理与分析》第4.6节“形态运算”。
EditMasks 纯手工编辑。应该尽量避免，如果手工操作不可避免，应该选用尽可能大的编辑单位，用编辑单位的边缘拟合目标边缘。之所以提出这一点，是因为根据迭代端点拟合（Iterative Endpoint Fitting）和最小均方误差曲线拟合（Mean Square Error, MSE）的思想，我认为在有足够的已知边缘点的情况下，手工拟合也是能够得到较为准确的边界的。参见附图4
Calculate3D 根据mask重建成三维模型（等值面绘制）。之所以把这个功能放在Segmentation菜单下，是因为有时候我们需要借助三维重建的结果反过来帮助分割，因为在重建的过程中可以对模型在三维空间上进行平滑等运算。如果不使用这个工具，一个可能的后果就是目标mask在冠面视图和矢向视图中投影的边界非常不平滑。
BooleanOperations对两个mask进行与运算、联合运算或者。不能直接求取mask，往往用于联合几种分割方法得到的mask，对于保证模型质量非常有用。

前面提过，要想得到完善的图像分割，最重要的是设计好分割的流程（著名的图像分割配准算法包ITK就是基于这样的理念）。我觉得在MIMICS中也应该继承这一思想。下面以我的一个胃部CT序列为例说明之（分割骨骼等硬组织一般比较简单）。在这个CT序列中，我们要分割胃壁（软组织）。一个基础性的步骤是灰度窗映射，也即在直方图控制处选择Soft Tissue，这样显示的时候MIMICS会让软组织的灰度区别更大，让其他灰度区别减小。虽然这带来了更大的噪声，但是也明晰了软组织内部的边界。如果没有设定合适的灰度窗，后续分割工作会极难进展。
基于上述分割原则，本文构建了一个分割流程。该分割流程，除了计算Polyline外，在冠向视图和矢向视图都加以应用，以获得最佳效果。即便分割非常精良，在三维重建的时候，仍然需要进行平滑和面片精简。

Some Techniques on Image Segmentation UsingMIMICS

MIMICS(Materialise Interactive Medical Image Control System) provides state-of-the-arttechnologies which bridges sliced image data and the mechanical analysis ofhuman tissue. Segmentation is a crucial part of the whole procedure because biomechanicalanalysis cannot be conducted on the basis of a wrongly segmented tissue.

Once the image slices are imported intoMIMICS, we treat it as volume data field. The purpose of segmentation is toextract a mask that represents the target tissue as accurate as possible, andfrom which we then reconstruct a 3D model. MIMICS gives several convenient toolswhich implements most common segmentation algorithms (except differentialoperator). Here are some introductions to these tools in the segmentation menu.

Thresholding is the most commonly-usedtool. It extracts voxels whose gray value is in a given range from image data. Thisis oftentimes the first step to separate an object from CT/MRI image.

Region Growing is a useful tool to detectconnectivity of a mask. Note that the result is calculated according to a givenmask and the location of seed point. The algorithm doesn’t take any directreferences form the gray values in the volume image data.

Dynamic Region Growing is a tool combinesthe functions of thresholding and Region Growing at the same time. Given a seedpoint, it returns a mask composed of neighbor voxels whose gray value is in agiven range. MIMICS official webpage says this is a convenient tool forsegmentation of nerve and blood vessel.

Morphology Operation includes dilate,erode, open and close operations. Although this tool does not make any use of thegray value data directly, it’s rather a Swiss army knife to smooth the edge ofa mask. Go to… if you want to get clearer of what this tool does.

Edit Mask is a convenient tool for manualsegmentation were we use our mouse to portray the mask. Because our mouse-clickinfluences the result mask, the use of this tool should be avoided. However, ifyou have to separate object tissues by hand anyway, you’d better use larger portrayunit. For example, in attached picture 4, assume the blue curve is the edgewe want to tell computer (it exists only in our mind at this time!) andeach circle represents the result of a mouse-click. Obviously, the largerradius each circle unit has, the smoother the result portrayed edge is. By usinglarge portray unit, we can reduce both the amount of manual work and theinfluence of inaccurate mouse-click. This technique was inspired by the thoughtof IEF(Iterative Endpoint Fitting) and MSE(Mean Square Error).


Calculate3D You may surprise that this isalso a tool that assists segmentation! We indeed sometimes get bettersegmentation result by virtue of the 3D reconstruction result ( using the Calculate Polylines form object tool). In theprocess of reconstruction, we actually smooth the mask in 3-D space. This is whatwe cannot achieve as we smooth the segmentation result slice by slice, or ineach 2-D image since some MIMICS segmentation algorithms and our manual workare performed two-dimensionally. If we don’t make use of this tool in our segmentation,a possible outcome would be that the mask edge looks perfectly smooth in axialview, but very coarse in saggital and coronal views, as shown in the attached picture 5.


Boolean Operations is like MorphologyOperations, which only operates the given mask instead of the image data. Operationson different segmentation results are useful to check errors in thesegmentation results. For example, if our target is a tube, like colon and we‘veget holes on the colon body in our segmentation result, now we invent a mask bydilating the inner space (with low gray value) by 1 or 2 pixels and then minus theinner space. This invented mask indicates the minimum thickness of colon. So wecan unite this mask and the segmentation result of other methods (with holes)and get a more reasonable result. This kind of method is a key to ensure segmentationquality.

Based on my theory, I made a segmentation procedure.See attached picture 1. Before getting into the procedure, remember to rescalethe image data to emphasize the gray values of the target tissue. Thisamplifies noises but clarifies edges at the same time.

[ 本帖最后由 jati 于 2008-4-12 10:49 编辑 ]

x3k

thanks for sharing!!!

walkmoon

图像分割是一个经典难题， 针对具体问题的不同应采取不同的分割策略，感谢楼主的分享，加分鼓励！

无疾

感谢分享！！！1

go95462357

解说的很透彻啊,楼主功底很深啊  赞

elneen

谢谢分享。

wonderful

很好很强大!

oceandefence

软组织怎么分割开来啊
区别很小

a0s0d0f0g0

很有用啊，感谢分享！

superants

谢谢楼主贡献，有用

iisltyf

真的很厉害

leeray918

楼主是用的12.1吗

tjfch

谢谢分享。利用膨胀和销蚀功能可以区分开皮质骨和松质骨吗？

walkmoon

不错!

hkdhhy

楼主 软组织怎么划分呀

hkdhhy

楼主好厉害。。。
看了步骤图 好多还是不懂
学习中

宋佳音24

谢谢楼主，学习中

chenjianming

求助：ct片如何使用mimics编辑??

419813694

楼主功夫非常深厚，不知mimics可否将不是同一系列的CT图片（例如，1是股骨的图片，2是胫骨的图片），同时导入mimics中形成膝关节？非常感谢

xiaoyuer

xiexielouzhu