Canlı vericili karaciğer nakli için karaciğer damar ağacı bölütleme

Pre-surgical evaluations of living donated liver transplantation require accurate segmentation of liver vasculature. Expert radiologists carry out this procedure either manually or using semi-automatic software tools. Manual delineation is very time consuming and tedious work and segmentation accuracy is largely dependent on the expert's abilities and very susceptible to human error. Semi-automatic methods are faster, however require advanced interaction mechanisms and iterative optimization. Thus, there is a need for automated methods. Unfortunately, the vascular tree of the liver is very complex and show high variability. Moreover, the contrast-enhanced images may contain significant amount of artifacts and task associated difficulties. Therefore, the development of a fully automatic method becomes a challenging task. Currently, there is no well-established datasets for comparative analysis of existing methods. This makes is it hard to propose improvements due to the lack of qualitative analysis of different techniques on a benchmark dataset. In this thesis, first, a database, which consists of 35 abdominal computed tomography angiography datasets, is collected and hepatic and portal veins are annotated manually. The process is carefully supervised by an experienced radiologist. Some of the well-known vessel segmentation methods were tested and their performances were analyzed. Finally, deep learning based methods were applied to reflect the performance of emerging deep models. After extensive experimentation, DeepMedic architecture is shown to achieve the best performance. An automatic system, which employs combinations of multi planar reconstructions, is developed. The obtained results are shown to outperform both the existing methods and the individual utilization of deep models.