Novo método de classificação automática de achados em mamografias FFDM e SFM utilizando rede neural

• Autor(a) principal: Melo, Matheus Cordeiro de
• Data de Publicação: 2017
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Biblioteca Digital de Teses e Dissertações da UFPB
• Texto Completo: https://repositorio.ufpb.br/jspui/handle/123456789/12927
• topic: Classificação de achados mamográficos; Ruído em mamogramas; Mamografias FFDM; Mamografias SFM; Classification of mammographic findings; Noise on mammograms; FFDM mammographies; SFM mammographies; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO

Breast cancer is the cancer with the highest incidence among women. Mammography is the most suitable exam for the early detection of this cancer. This exam detects small lesions and it allows the visualization of up to 90% of the abnormalities (masses, calcifications, architectural distortions). The diagnosis of breast cancer is a process prone to errors due mainly to doctor’s misinterpretations, consequence of the real difficulty of the procedure. Computer Aided Diagnostic (CAD) systems are developed with the aim of being a second opinion in the identification and diagnosis of the disease using mammographic images. Mammographic images can be obtained through Screenfilm Mammography (SFM) and Full Fiel Digital Mammography (FFDM). FFDM presents a better resolution than SFM and is gradually replacing the SFM. This work developed a new method for automatic classification of mammographic findings in benign or malignant. Firstly, two databases were selected, the Digital Database for Screening Mammography (DDSM) that offers images from SFM and INbreast that has images obtainedbyFFDM. Next, images withmammographicmasses from DDSMand from INbreast were selected randomly. In the pre-processing stage, a new technique of noise identification and elimination based on Information Theory was developed. After the pre-processing, important descriptors were extracted through the Gray-Level Co-Occurrence Matrix and the method Segmentation-based Fractal Texture Analysis. Finally, Finally, the base of characteristics served as entry forthe multilayer perceptron neural network classifier. Considering DDSM, the method reached 93.13% of accuracy, 94.17% of specificity and 91.23% of sensitivity. Using INbreast, 88.37% of accuracy, 86.49% of specificity and 89.80% of sensitivity were achieved. In this work the obtained results proved to be competitive with the literature of the area.