Método de superfícies ativas usando local binary patterns (LBP) aplicado na segmentação de lobos pulmonares em imagens de tomografia computadorizada do tórax

Abstract

In several applications involving medical image analysis, the process of image segmentation, be it automatic or manual, is a present task. An accurate segmentation provides information for inspection of anatomical structures, to identify diseases and monitoring of its progress, and even for surgical planning and simulation. Thus, the role of image segmentation is essential in any medical image analysis system. Among the segmentation techniques in the literature, the active models technique is one of the most popular approaches of the last two decades and has been widely used in medical image segmentation, achieving considerable success. Active models that are applied on three-dimensional applications are called Active Surfaces Methods (ASM), which has been widely used in the segmentation of 3D objects, evolving under the influence of their energy to converge to the desired surface. So, knowing how essential surface extraction is to obtain an accurate segmentation, this thesis conducts a study on ASM, identifying its advantages and limitations, and proposes a new ASM to the segmentation of pulmonary lobes on CT images. The new ASM has as contributions internal forces for unstructured meshes, external energies based on LBP texture and Hessian matrix, and automatic initialization for each lobe. In order to validate this proposal, a comparative study of the performance of the internal forces in synthetic images, along with the comparison of the segmentation of lung lobes obtained by the proposed method with the segmentation of a gold standard carried out by an expert medical board will be conducted. The results show that the internal forces performs well, providing synthetic images segmentation with average distance of less than 1 voxel and adjustment measures of 0.95. The automatic initialization has also achieved significant results, with overall hit rate of 94%. Finally, the rates obtained for pulmonary lobe segmentation allows validation of the proposed method with average distance values of 1.93 mm and rates of size and form adjustment of 0.98 and 0.89, respectively. Thus, it is concluded that the obtained metric is sufficient to validate the lobar segmentation obtained in this thesis