Detector de pontos salientes 3D baseado na DT-CWT com aplicação no posicionamento de malhas deformáveis em imagens de ressonância magnética do cérebro

• Ano de defesa: 2019
• Autor(a) principal: Souza, Breno da Silveira
• Orientador(a): Ferrari, Ricardo José
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de São Carlos; Câmpus São Carlos
• Programa de Pós-Graduação: Programa de Pós-Graduação em Ciência da Computação - PPGCC
• Palavras-chave em Português: Detector de pontos salientes 3D; Transformada wavelet complexa de árvore dupla; Modelos geométricos deformáveis; Imagem de ressonância magnética
• Palavras-chave em Inglês: Salient point 3D detector; Dual-tree complex wavelet transform; Geometric deformable models; Magnetic ressonance image
• Área do conhecimento CNPq: CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::SISTEMAS DE COMPUTACAO
• Link de acesso: https://repositorio.ufscar.br/handle/20.500.14289/12071

The magnetic resonance (MR) imaging is a noninvasive method that allows high soft tissue differentiation by providing a good anatomical detailing level. For this reason, the MR image is often used to help in diagnosis and supervision of neurodegenerative diseases, in which we can mention the Alzheimer's disease (AD). The analysis of shape changes and gradual volume reduction of brain structures of interest helps to identify the AD staging. Considering that the manual annotation of those structures in the MR image (to posterior volume and shape evaluation) is highly susceptible to measurement error, several automatic methods have been proposed in the literature for this task. Some of the most popular among the proposed methods are so-called geometric deformable models. They are able to catch the geometric variability of anatomical structures because they offer a priori knowledge of shape and localization. However, despite these advantages, geometric deformable models need to be positioned close to the strutcture to be segmented to achieve success. One way to minimize this problem is to use salient points to estimate a local deformable transformation, which will be applied to the mesh vertices of a given structure to perform its initial positioning. In this project, we propose a new approach for the detection of 3-D salient points in MR images which is based on the dual-tree complex wavelet transform (DT-CWT). Our method combines filter response maps calculated for different scales and orientations of the DT-CWT to create an accumulated map of energy responses. We tested our approach for the positioning of hippocampi meshes in 3-D brain MR images and compared the results with manual annotations made by a medical specialist. Mean values of Dice (DC) and Hausdorff Average Distance (HAD) measurements of our proposed method showed better results (DC = 0.58/0.58 and HAD = 0.73/0.75 for the left and right hippocampus, respectively) when compared to both an Affine transform guided initialization (DC = 0.49/0.50, HAD = 1.04/1.11) and a transform guided initialization using salient points detected by a phase congruency technique (DC = 0.55/0.56, HAD = 0.84/0.83).