|
www.univ-lyon2.fr/recherche/actualites/soutenance-de-these-jonas-lamy ; Hepatic vessels segmentation in medical images is still an open problem. Yet, such segmentation is useful for praticians in a couple of contexts including surgical planning. The difficulties of hepatic vessels segmentation are numerous. Compared to the volume of the liver, the vessels cover a sparser area. They exhibit a weaker contrast compared to neighbouring tissues, a variable morphology and for the smallest vessels, a size nearing the images resolution. The imaging context of computed tomography (CT) and magnetic resonance imaging (MRI) adds an extra layer of difficulties induced by specific artifacts (noise, contrast inhomogeneity, movement, etc.). To ease the segmentation process, one can use vesselness filters. These filters can improve the contrast of vessels by enhancing tubular structures. A number of these filters has been proposed in the litterature, however the usage of the same filters with default parameters has been observed in the majority of cases. In this context, our objectives have been to understand this trend, to compare the performances of several filters and to study their effectiveness in the context of liver vessels imaging. We chose to build a reproducible and extensible benchmark able to evaluate the filters over areas of interest defined by the user. We then put this benchmark in practice by analysing seven vesselness filters over tree datasets (CT scans, MRI, and synthetic MRI). An exhaustive analysis of the filters is performed over six areas of interest : the organ, the vessels neighbourhood by size and the bifurcations. We also discuss the effect of the parameters of the filters and give usage recommendations depending on the segmentation context. All the code needed for our experiments is publicly available : the annotations tools, datasets and ground truth as well as filters implementation and an online demonstrator. This PhD thus regroups a comprehensive set of knowledge and tools for anyone ...
|
