Detecção de nódulos em imagens mamográficas utilizando aprendizado profundo e transformada Wavelet
Data
Autores
Título da Revista
ISSN da Revista
Título de Volume
Resumo
O câncer de mama permanece como uma das principais causas de mortalidade entre mu- lheres no mundo, sendo a detecção precoce essencial para aumentar as chances de trata- mento eficaz. Nesse contexto, este trabalho propõe o desenvolvimento e a comparação de algoritmos para detecção automática de nódulos mamários em mamografias, integrando técnicas de pré-processamento baseadas na Transformada Wavelet e redes neurais con- volucionais (CNNs), treinadas tanto do zero quanto com Transfer Learning. As imagens utilizadas provêm da base pública Mini-MIAS. O desempenho dos modelos é avaliado em diferentes cenários, com e sem a aplicação da Wavelet e com o uso de data augmentation, utilizando métricas como acurácia, precisão, recall e F1-score. Os resultados mostram que o pré-processamento baseado na família Coiflet exerce maior impacto no desempenho do que a data augmentation isolada, reduzindo o desequilíbrio entre as classes e melhorando a identificação de padrões relevantes. Entre as arquiteturas avaliadas, os modelos ResNet apresentaram desempenho mais consistente, com destaque para a ResNet34 combinada com o pré-processamento Coiflet, que obteve os melhores resultados gerais.
Breast cancer remains one of the leading causes of mortality among women worldwide, and early detection is essential to improving the chances of effective treatment. In this context, this work proposes the development and comparison of algorithms for the automatic detection of breast nodules in mammograms, integrating preprocessing techniques based on the Wavelet Transform and convolutional neural networks (CNNs), trained both from scratch and using Transfer Learning. The images used come from the public Mini-MIAS database. Model performance is evaluated under different scenarios, with and without the application of the Wavelet transform and with the use of data augmentation, using metrics such as accuracy, precision, recall, and F1-score. The results show that preprocessing based on the Coiflet family has a greater impact on performance than data augmentation alone, reducing class imbalance and improving the identification of relevant patterns. Among the evaluated architectures, the ResNet models presented the most consistent performance, with ResNet34 combined with Coiflet-based preprocessing achieving the best overall results.
Breast cancer remains one of the leading causes of mortality among women worldwide, and early detection is essential to improving the chances of effective treatment. In this context, this work proposes the development and comparison of algorithms for the automatic detection of breast nodules in mammograms, integrating preprocessing techniques based on the Wavelet Transform and convolutional neural networks (CNNs), trained both from scratch and using Transfer Learning. The images used come from the public Mini-MIAS database. Model performance is evaluated under different scenarios, with and without the application of the Wavelet transform and with the use of data augmentation, using metrics such as accuracy, precision, recall, and F1-score. The results show that preprocessing based on the Coiflet family has a greater impact on performance than data augmentation alone, reducing class imbalance and improving the identification of relevant patterns. Among the evaluated architectures, the ResNet models presented the most consistent performance, with ResNet34 combined with Coiflet-based preprocessing achieving the best overall results.
Descrição
Citação
CARRICO, Jéssica Gomes. Detecção de nódulos em imagens mamográficas utilizando aprendizado profundo e transformada Wavelet. 2025. Trabalho de Conclusão de Curso (Bacharelado em Engenharia de Telecomunicações) - Instituto Federal de Educação, Ciência e Tecnologia de Santa Catarina, São José, 2025.
