Research On Ultrasound Image Segmentation Method In The Non-invasive Therapy Of Tumor Using Focused Ultrasound

The technology of using high-intensity focused ultrasound (HIFU) for the minimally invasive or non-invasive therapy of tumor has attracted extensive attention in medicine over the past few decades. The therapy mechanism of HIFU is that the ultrasonic wave can penetrate human skin and be focused on the target diseased tissues in the body under the guidance of medical imaging to generate the instantaneous high energy to damage the target tumor directly. The technology of medical imaging has played an irreplaceable key role in providing important diagnostic information for the location and measurement of the diseased tissues, the pre-operation treatment planning, real-time monitoring and guidance in the operation, the postoperative treatment of assessment. At present, there are mainly two kind of image guidance system, based on MRI image-guided HIFU therapy system and based on ultrasound image-guided HIFU therapy system. Due to the cheap, real-time characteristics of ultrasound imaging, the ultrasound image-guided HIFU therapy system is widely used in clinical research and application. However, because of the mechanism of ultrasound imaging, it universally exists a lot of speckle noise in ultrasound images, leading to a low resolution and low contrast of ultrasound images, badly blurred edge of the tissues, which make it difficult to detect the edge of the target tumor and segment accurately. In clinical application, the extraction of diseased tissues is obtained by manual drawing the contour in the ultrasound images by a skilled doctor, which is a time-consuming heavy workload. Ultrasound image processing technology lag behind in ultrasound imaging, which limits the application of ultrasound images in medicine. Introducing the method of automatic feature extraction and analysis in the ultrasound image processing has important significance. This paper researches on the two key issues of noise reduction and segmentation in ultrasound image processing, including the following several aspects:1) In noise reduction of ultrasound images, after detailed research into the method based on the traditional anisotropic diffusion filter for speckle noise suppression this paper proposes an improved anisotropic diffusion filter. A scale factor is introduced to describe the relationship between the instantaneous coefficient and the speckle scale function, which are the two key components of the diffusion coefficient in the traditional anisotropic diffusion filter. The scale factor is obtained as the OTSU segment thresh of the edge indicator function of the Gaussian filtered image. The scale factor is used to replace the speckle scale function in the diffusion coefficient. The best advantage of the scale factor is to avoid the difficult problem that the filter parameters in the speckle scale function are hard to set. In another aspect, the diffusion process is terminated by giving the number of iterations in advance. This paper proposes a way to stop the diffusion process automatically. The speckle signal to noise ratio of the filtered image after each iteration is monitored to calculate the relative changes. When the relative change is smaller than a given threshold value, the diffusion process terminate. The proposed anisotropic diffusion filter doesn't need to set filter parameters and can terminate automatically. The simulation experiments and experiments of the real ultrasound images show that the proposed anisotropic diffusion filter can be used in image processing more practically and conveniently.2) In segmentation of ultrasound images, firstly, the parameter active contour model is detailed studied. The mathematical principles and the segmentation performance of three classic external force active contour models, which are the traditional Snake, the balloon Snake and the gradient vector flow (GVF) Snake, are compared and analyzed. The balloon Snake model is sensitive to the position of the initial contour, which means that the initial contour must be delineated completely in the interior or the exterior of the real boundary of the target in ultrasound image because the direction of the balloon force is set as the inward or outward normal vector direction of the evolution curve at the same time. In order to solve the problem, this paper proposes an adaptive balloon Snake model. The direction of balloon force is determined by two vectors, which are the boundary vector obtained from an edge map and the outward normal vector of the evolution curve. An improved coefficient variation edge map by OTSU method is used in the proposed balloon Snake model to reduce the interference of speckle noise. The segmentation results of the ultrasound image sequences show good performance of the proposed balloon Snake model.3) In segmentation of ultrasound images, secondly, the traditional level set active contour models are detailed studied. The mathematical principles and the segmentation performance of three classical level set active contour models, i.e., geometric active contour model, geodesic active contour model and Chan-Vese model, are compared and analyzed. The geodesic active contour model is a kind of level set active contour model based on image edge information, it considers only the local information of image edge in the process of image segmentation, makes the contour evolution into a local minimum value easily. Chan-Vese model is a kind of level set active contour model based on image region information, it considers the global information of the whole image in the process of image segmentation, makes wrong region segmentation of the low contrast image easily. In order to solve these problems, this paper proposes an combined energy model. Under the combination with the edge information and region information in image, the proposed level set active contour combines the advantages of geodesic active contour model and Chan-Vese model and overcome the defects of the two models. The experiment results show that the proposed model has good performance in segmentation with weak edges of low contrast gray image and has good global optimal segmentation performance.