Classification of breast injuries in categories 4 and 5 of the BI-RADS® standard using neural networks

• Autor(a) principal: Nery Júnior, Elmo de Jesus
• Data de Publicação: 2022
• Outros Autores: Silva Neto, Otílio Paulo da, Ribeiro, Francisco Adelton Alves, Lima, Francisco das Chagas Alves, Verdes, Larysse Maira Cardoso Campos, Silva, Danylo Rafhael Costa, Oliveira, Maria da Conceição Barros, Diniz, Pedro Henrique Bandeira, Paiva, Anselmo Cardoso de, Silva, Aristófanes Corrêa
• Tipo de documento: Artigo
• Idioma: por
• Título da fonte: Research, Society and Development
• Texto Completo: https://rsdjournal.org/index.php/rsd/article/view/31305
• topic: Breast cancer; BI-RADS® Classification; Image processing; Neural networks.; Cáncer de mama; Clasificación BI-RADS®; Procesamiento de imágenes; Redes neuronales.; Câncer de mama; Classificação BI-RADS®; Processamento de imagem; Redes neurais.

Breast cancer is the disease with the highest incidence among women worldwide, with an estimate for Brazil in the 2020-2021 biennium about 66,280 new cases of breast cancer, which corresponds to a rate of 29.7% of cases in the female population and about 15,000 deaths from the disease. Mammography is one of the most used tests for early detection of this type of neoplasm. However, errors occur in the reading and interpretation of reports, even a well-trained professional has a success rate between 65% and 75% with an amount of false negative varying between 15% to 30% and a false positive of 7% to 10%, resulting in an unnecessary amount of biopsy, 65% to 90% of tissue biopsies with suspected cancer are benign, causing emotional and physical repercussions for patients. Computer systems can be developed to aid in medical diagnosis. This article applied neural network techniques to develop a computational tool capable of classifying injuries from categories 4 and 5 of the BI-RADS® standard. The results acquired by the software, observed that the best classifier with regard to the accuracy rate was Deep Learning, reaching a percentage of 82.60%, the Support Vectors Machine - SVM had a percentage of 73.97%. This demonstrates that the neural network techniques used in the software design show an efficiency in the lesion classification task.