Um Novo Sistema Automático para Detecção e Classificação de Nódulos Pulmonares em Imagens de Tomografia Computadorizada do Tórax Usando uma Única Rede Neural Convolucional

• Ano de defesa: 2021
• Autor(a) principal: Ribeiro, Alyson Bezerra Nogueira
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Palavras-chave em Português: TC do Tórax; Rede neural convolucional; Detecção de nódulos pulmonares; Visão computacional
• Link de acesso: http://www.repositorio.ufc.br/handle/riufc/61105

The Computer Vision use to aid diagnosis in imaging exams has a great advancesis in scientific community. Due this context, lung cancer identification and treatment research stands out. Considered one of biggest world preventable causes of death, this cancer type is included in 10 diseases group of most deaths cause in Brazil. Numerous studies are carried out to provide a detection and analysis of lung nodules using Computer Vision algorithms and other methodologies. For nodule detection, convolutional neural networks application stands out among strategies. On the other hand, they are also applied image processing techniques in literature to segmento nodules and attributes extaction. The Hessian matrix use is recurrent as analysis factor for nodule segmentation and nodule candidate detection algorithms. These surveys are generally based on very specific conditions or dependent on preliminary step executions. In this case, gold standard works contained lung segmentation and candidate selection to generate nodule detection results. In this context, this thesis propose a new methodology for nodules detection, classification and a data external independency. It is also a complete framework divided into three steps: the first use the 3D image curvedness combined with Otsu thresholding for candidate detection, the second is a convolutional neural network application using a new architecture for candidate classification and, finally, texture, malignancy, contour margin and size of lung nodules classification in thorax CT images. The nodules detection results has a 0.902 sensitivity for a 8 false positives per exam. In case of attribute classification, the malignancy classification has sensitivity 0.940.