Study On SVM And Its Application In Image System Of Phased HIFU

In recent years, phased High Intensity Focused Ultrasound (HIFU) therapy technique has become a new hotspot in the research field of tumor therapy and therapeutic ultrasound. Using the ultrasound peculiar capability of penetration, orientation and focusing, phased HIFU can build high intensity focus in deep tissue and cause acute cell death by focusing ultrasound beams without damaging the close normal tissue. And it has become an efficient method of the minimal-invasive therapy of tumor. To protect normal tissues and increase the treatment efficiency, the accurate location of the tumor target must be provided. It depends on the image guided system of HIFU to acquire the postion information of tumor target because HIFU surgy is non-invasive. It is the key factor of a successful HIFU therapy. In order to solve the problems in the image guided system such as time consuming, low precision and bad robustness, this paper study on support vector machine (SVM) and its application in the image system, and the results are satisfied.The SVM approach is considered as a good candidate for utilizing the prior knowledge because of its high generalization performance and sparse solution. SVM is a new type of learning machines based on statistical learning theory. It follows the principle of structural risk minimization, not the traditional principle of empirical risk minimization. The HIFU image system can be improved by using SVM correctly. In this paper, an overview on the academic base of SVM-- statistical learning theory is given. SVM's applications in pattern recognization and regression estimation are introduced. A density estimation method based on linear SVM is given. Furthermore, a method for density estimation is developed based on the Multi-Kernel SVM. We extend these methods into the Multi-dimensional density estimation problem.In the image guided system of phased HIFU we researched includes two main parts: image segmentation and image registration. In order to make therapy plan before operating, CT or MRI images are segmented. Then the reasonable therapy path of the HIFU focusing beam is built according to the result of segmentation. And image registration is to integrate the pre-operative image with the intra-operative image, to transform the pre-operative coordinates to the operating coordinates, to map the therapy plan including tumor target contour and treatment paths to the operating coordinates. In this background, this paper study on some new medical image segmentation and registration methods based on SVM. For image segmentation, two methods based on SVM for density estimation are presented. All of them improved level set method by incorporating prior knowledge into the curve evolution. One method used SVM to construct a prior model about the image intensity and curvature profile of the structure from training images. When segmenting a novel image being similar to the training images, the technique of narrow level set method is used. The higher dimensional surface evolution metric is defined by the prior model instead of by energy minimization function. The other method used SVM to construct a prior model of the shape of the desired object from training images. When segmenting a novel image, we improved level set method based on C-V model by incorporating this prior model. At each step of the curve evolution, we estimate the maximum a posteriori (MAP) shape of the object according to prior shape model. And then we evolve the curve towards the MAP estimate. Segmentation results are demonstrated on synthetic images, MR images and ultrasonic images. It shows that the prior knowledge model makes segmentation process more robust and faster. For image registration, a method of computing different modality medical images registration of the same patient is presented. It incorporates prior joint intensity distribution between the two imaging modalities based on registered training images. The prior joint intensity distribution is modeled by support vector machine. Results aligning CT/MR and Pet/MR scans demonstrate that it can attain sub-voxel registration accuracy. Furthermore, it is a fast registration method because support vector machine solution is sparse.