http://chineseinput.net/에서 pinyin(병음)방식으로 중국어를 변환할 수 있습니다.
변환된 중국어를 복사하여 사용하시면 됩니다.
Improved variational methods in medical image enhancement and segmentation
Phan Tran Ho Truc 경희대학교 2009 국내박사
Medical procedures have become a critical application area that makes substantial use of image processing. Medical image processing tasks mainly deal with image enhancement and segmentation, i.e., filtering, sharpening, and bringing out image details, and quantitatively measuring medical conditions such as vessel size, tumor volume, or bone fracture length. Over a long time, medical image processing problems were tackled by a pool of heuristic and low level mathematical operators. Simple heuristic approaches, e.g., histogram equalization, can provide surprisingly good results; however, one cannot know exactly when and why they work or do not work. Recently, the use of variational methods in image processing has emerged as an interesting research topic. Variational methods can be based on either the optimization of a energy functional or the design of a partial differential equation (PDE) whose steady state is the solution of the problem at hand. The advantage of these methods is that the theory behind the concept is well defined. However, the direct application of variational method (as it is) to medical image processing suffers from limited performance since medical images are of very poor quality. It means that purely image-driven methods can hardly work for medical images. In order to find a better solution, one has search on a solution space that is constrained by prior knowledge about the object of interest. In this thesis, I proposed two novel variational algorithms for medical image processing that incorporate knowledge about the structure of interest (vessels, tissues, or bones, etc.) and the imaging modality which is usually known in advance. The first algorithm involved the design of a PDE for curvilinear structure enhancement. Curvilinear structures are targeted as they resemble the image representation of an important organ: blood vessel, which is critical in diagnosis of many injuries in heart, retina, or brain, etc. Here, the prior knowledge is the fact that a vessel has linear elongated structure and a Gaussian proflle across its width (the intensity is highest at the vessel's center and gradually decreased towards its two sides). This is because the blood itself or the contrast agent it carries determines the contrast of the vessel, leading to higher intensity at the vessel's center and lower intensity at its sides. Unlike conventional approaches that use Hessian tensor to detect the elongated structure and suffer from the junction suppression problem, the proposed approach utilizes directional filter bank (DFB). DFB provides more global directional information than the Hessian tensor does. The proposed approach was shown to be able to overcome the junction suppression problem. The second algorithm involved the development of an active contour for inhomogeneous structure segmentation. Here, the prior knowledge resides in the fact that distinct organs shall generate distinct configurations (intensity) in the image. As a result, density function of the foreground (object of interest) should be at far distance from that of the background. Chan-Vese model expresses this "difference" characteristic by only the mean values of the foreground and the background, which is effective for largely homogeneous structures. However, objects in medical images, e.g., bones in CT images, are often inhomogeneous. The proposed model therefore expresses the "difference" characteristic using the whole density function itself and reflects it to the energy functional in the form of a Bhattacharrya distance. Minimization of the proposed energy functional leads to a novel active contour model that can robustly segment inhomogeneous objects. The two proposed models have been extensively evaluated using various synthetic images as well as real medical images such as angiography and CT in comparison with many existing approaches. Experimental results showed that the proposed models provide better performances most of the time.