3D AUTOCUT: método de segmentação 3D aplicado a imagens de tomografia computadorizada do tórax

• Ano de defesa: 2018
• Autor(a) principal: Cavalcanti Neto, Edson
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Palavras-chave em Português: Teleinformática; Processamento de imagens - Técnicas digitais; 3D Segmentation; Automata
• Link de acesso: http://www.repositorio.ufc.br/handle/riufc/42010

Among all types of cancer, lung cancer (LC) is one of the most common of all malignancies, presenting a growth of 2% in its worldwide expansion. In Brazil, for the year 2018, 31,270 new LC cases are estimated, of which 18,740 are in men and 12,530 in women. This is important to prevent the diagnosis and increase the chances of cure. In such diseases, the Computed Tomography (CT) scan of the Thorax contains information from the 3D segmentation of the lungs helping its early diagnosis. Different methods have been developed for the 3D lung segmentation, however, they mostly present problems when there are diseases such as LC or other structures present inside the lungs. In this thesis, a new 3D algorithm is proposed, which uses a force composition in the RGB space and then initializes the labels and seeds that evolve interactively until they stabilize. In addition, a new 3D automatic initialization method for lung segmentation in CT images is also proposed. The tests were performed for 3D segmentation in synthetic images (cylinder, doublecone, cube and sphere) and images CT scans of the thorax. In the tests of the synthetic images are introduced noises to evaluate, based on the position adjustment and the Dice similarity coefficient, the 3D segmentation capacity and robustness of the method, compared to other methods found in the researched literature. In this type of images, the results obtained showed that the 3D Autocut algorithm has the best results for both low noise and high noise images with position adjustment measures and Dice coefficient above 0.95. The tests for thoracic CT images, based on the metrics Dice similarity coefficient, position adjustment, shape and size adjustments, use 15 exams in the apex, hilo and base positions of healthy and pathological patients. The 3D segmentation result, based on the 3D Autocut method, as well as the use of the proposed initialization, also produces superior results in 93.3% of the exams when comparing the final form of segmentation and Dice Similarity. The main contributions are the Autocut 3D algorithms and the automatic seed initialization algorithm.