应用蠕虫模型自动分割医学图像的研究

A Worm Model for Automatic Segmentation of Medical Image

目的将人工生命的原理应用于图像自动分割,提高图像分割的准确性。方法构造了一种具有初步生命特征的可变形模型——蠕虫模型。该模型由4种参数进行约束,是由若干节点组成的中线结构。结果该蠕虫模型可以将胼胝体从二维MR矢状位图像中自动地完整地分割出来。结论该模型能综合利用多种信息进行图像分割,具有常规图像分割方法不具备的能力。对于边界有间断的较复杂的图像,本模型的优势更加明显。

Objective The correct classification of human organs in medical image can provide computer -assisted diagnosis for clinicians, and also it is the groundwork of some image processing, such as the 3D reconstruction and the visualization of medical images. There are many segmentation algorithms by far, such as threshold segmentation method, differential coefficient operator edge detection method, region growing method and clustering segmentation method, however, they are far from perfect. Most of them are semiautomatic algorithms which can not receive satisfied segmentation result without the intervention of experts. Medical image segmentation has always been the research hotspot in the medical image analysis area, and is also the key step to finish the quantitive analysis of the anatomical structure and extract the related diagnosis information. The variety properties of image as well as the mixture with a lot of noise lead to the main difficulties in the image segmentation. In this article we try to increase the accuracy of segmentation by applying the principle of artificial life to automatic segmentation of images. Methods The method based on artificial life is useful in understanding biological rides, and it also has been successfully applied in the robot and computer graphics. It is a new trial to apply the principle of artificial life to medical image segmentation. We constructed an intelligent deformable model called worm model with the application of image segmentation based on artificial life and created a worm model. The worm model, no larger than several pixels, is made up of many nodes, and it possesses ‘ head’ ,‘ neck’ and ‘ body’. The ‘ head’ is the function section, which has a nerve center system, vision system and perception system. It also can memorize information and give off orders. The ‘ neck’ and the ‘ body’ are locomotion parts. The worm has a central nervous system, vision system, perception system and locomotion system. It can memorize, recognize objects and control the motion of its body. Firstly we converted a MR image into a binary image, and then a worm model was placed on the top of this image. Using the prior experience, the worm would automatically find the corpus callosum. Then it would go down to the CC ＇ s center where it would divide into four little worms in order to detect edges in different directions. We commanded these worms detect image edges in four different directions： up-left, up-right, bottom-left, and bottom-right. At last, the detecting tracks were memorized when the image edge was detected completely. Results The worm model overcomes the defects of the current existing methods since it is able to intelligently process the segmentation of the image using more available information, rather than only using pixels and gradients. Its ability to segment the CC from MRI brain images showed that the worm model could segment images automatically and accurately. For those images that are more complex or with fragmentary boundary, the predominance of the worm model is especially clear. Conclusion The model is able to intelligently process the segmentation of images using more comprehensive information rather than using pixels and gradients only. The worm＇ s biological activities are analyzed deeply and it is proved that the worm model can be used in image segmentation. This model has nerve center system, vision system, perception system and motor system, and it also has some life characteristic, such as memorizing, cognizing and locomotion controlling. Because the model has alterable eyeshot, as well as can deform flexibly , it is able to take full advantage of local information . As a result , it can automatically segment the target using more information than just pixels and edges, and then it can overcome the defects of the traditional algorithms. Further more, it has a high intelligence and only need a few prior experiences. The experiment result based on simulation data and real data proved the validity of the method.