Detecção de regiões de massa por análise bilateral adaptada à densidade da mama utilizando índices de similaridade e redes neurais convolucionais

• Ano de defesa: 2017
• Autor(a) principal: DINIZ , João Otávio Bandeira
• Orientador(a): SILVA, Aristófanes Corrêa
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Maranhão
• Programa de Pós-Graduação: PROGRAMA DE PÓS-GRADUAÇÃO EM CIÊNCIA DA COMPUTAÇÃO/CCET
• Departamento: DEPARTAMENTO DE INFORMÁTICA/CCET
• País: Brasil
• Palavras-chave em Português: Imagens médicas; Registro de imagens; Índices de similaridade; Redução de falsos positivos; Redes neurais convolucioanis.
• Palavras-chave em Inglês: Medical images; Image registration; Indexes of similarity; Reduction of false positives; Convolucioanis neural networks
• Área do conhecimento CNPq: Processamento Gráfico; Cancerologia
• Link de acesso: http://tedebc.ufma.br:8080/jspui/handle/tede/1537

Breast cancer is the type of cancer that most affects women and is one of the leading causes of death worldwide. Aiming to aid the detection and diagnosis of this pathology, several techniques in the image area are being created serving as a second opinion. It is known that mammograms of the left and right breast present a high degree of symmetry, and when there is a sudden difference between the pairs, it can be considered suspicious. It is also emphasized that the breast can present different density of the tissue and this can be a factor that makes difficult the detection and diagnosis of the lesions. Thus, the objective of this work is to develop an automatic methodology for the detection of mass regions in pairs of digitized mammograms adapted to breast density, using image processing and species comparison techniques to determine asymmetric regions in the breasts together with neural convolutional networks for Classification of breast density and regions in masses and not masses. The proposed methodology is divided into two phases: training phase and test phase. In the training phase will be created three models using convolutional neural networks, the first able to classify the breast as density and the last two to classify regions of mass and non-mass in dense and non-dense breasts.The steps are in aligning the breasts so that it is possible to make a comparison between the pairs. When comparing, asymmetric regions will be segmented, these regions will undergo a process of reduction of false positives in order to eliminate regions that are not masses. Before classifying the remaining regions, the breasts undergo the process of density classification by the model obtained in the training phase. Finally, for each type of breast, a model will classify the regions segmented into masses and not masses. The methodology presented excellent results, in the non-dense breasts reaching sensitivity of 91.56 %, specificity of 90.73 %, accuracy of 91.04 % and rate of 0.058 false positives per image. Dense breasts showed 90.36 % sensitivity, 96.35 % specificity, 94.84 % accuracy and 0.027 false positives per image. The results show that the methodology is promising and can be used to compose a CAD system, serving as a second option for the expert in the task of detecting mass regions.