A Minimally Supervised Classification Method Based On Statistical And Spatial Information And Its Application In Liver Segmentation

In this thesis a minimally supervised classification method using both statistical and spatial information is developed and applied on a CT liver segmentation problem. The method woks on the enhanced CT abdomen data and produces the segmentation results of liver, portal vein and hepatic vein. The results have been checked by the doctor and been regarded as clinical correct. This application is quite useful for the hepatectomy, especially the live donor liver transplantation.The live donor liver transplantation cuts off a portion of a healthy liver from a donor and replaces the entire liver of a patient who has a severe liver disease with it. It is the most difficult operation among all kinds of liver transplantations. The doctors must compute the volume of the portion to be cut off very carefully and accurately to ensure the safety of both the donor and the donee. The transplantation includes the division and anastomoses of the hepatic artery, portal vein and hepatic vein, which demands quite a complex and precise operation, so that the doctors need to observe the CT data repeatedly before the operation in order to be convinced of the shape of the liver and the distribution of vessel inside. The method introduced in this thesis can help the doctor to do such work and to plan the surgery by providing the segmentation results of liver, portal vein and hepatic vein, which can be easily reconstructed as 3D models.The two innovation points of the thesis are described below:First, from the view point of a segmentation method, the method in this thesis effectively overcome the problems caused by image noises and some other tissues, such as the heart, the intestine, the stomach, the muscle and the aorta abdominalis, which have similar CT values (H.U.) and are in close proximity (sometimes even connected) to the liver, through using both the statistical and spatial information skillfully and introducing the concept of high-confidence points and intensity-based distances. In addition, the method takes the advantage of the intensity differences among the three periods of the enhanced CT scan to segment the portal vein and hepatic vein, the results of which show the branches quite clearly until the fourth level, which sufficiently fit the doctor's requirement.Second, from the view of a clinical application, the segmentation of the portal vein and hepatic vein is definitely important and valuable for the planning of the live donor liver transplantation, but few efforts has been made internationally. Currently, the maximum intensity projection method is the most widely used method for the doctors to observe the structures of portal vein and hepatic vein. Since the method provides only pseudo-3D (or actually 2D) information, doctors have to use their knowledge and experience to imagine the real structure of the vessel in 3D space. However, the segmentation results of the method in this thesis can easily be reconstructed as 3D models and gives precise 3D information of vessels for doctors to plan the surgery more effectively and efficiently.