Desenvolvimento de rede neural convolucional utilizando classificadores para predição da agressividade biológica de subtipos de câncer de mama a partir de imagens de ressonância magnética de pacientes de hospital de referência do interior do estado de São Paulo
Carregando...
Arquivos
Data
2023-02-24
Autores
Esposto, Fábio Augusto da Silva
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Estadual Paulista (Unesp)
Resumo
Introdução: O Câncer de Mama é uma doença complexa e heterogênea, atualmente classificada em subtipos específicos que estão associados a diferentes prognósticos. A biópsia continua sendo o padrão-ouro para identificação dos biomarcadores, mas tem importantes limitações. A utilização de Aprendizagem de Máquina (AM) não é uma novidade na pesquisa do câncer. Redes neurais artificiais (RNAs) têm sido usadas na detecção e diagnóstico de câncer há aproximadamente 20 anos. Objetivos: Assim, essa pesquisa tem como objetivo desenvolver uma rede neural convolucional (RNC) que possa prever o subtipo específicos de um câncer de mama, classificados pela imuno-histoquímica, com base em imagens obtidas por ressonância magnética. Métodos: O estudo foi realizado no Hospital das Clínicas da Faculdade de Medicina de Botucatu em exames de ressonância magnética de 76 pacientes de gênero feminino. Os cortes das imagens de ressonância magnética foram usados para segmentação dos subtipos tumorais. Uma arquitetura de RNC foi projetada com uma máquina de Boltzmann convolucional personalizada utilizando exame de ressonância magnética de 76 pacientes, contendo 10 imagens cada, que foram utilizadas para treinamento e validação da RNC. As imagens foram ajustadas para validação. Um conjunto balanceado de 40 imagens (exame de 4 pacientes) foram utilizadas para definição de testes, sendo imagens de uma paciente para cada subtipo tumoral (Luminal A, Luminal B, HER2+ e Triplo Negativo). O código do software foi escrito em Python em uma estação de trabalho pessoal. Resultados e discussão: Dez imagens de cada subtipo tumoral foram avaliados (Luminal A, Luminal B, HER2+ e triplo negativo). Foram analisados o desempenho de cinco métodos classificadores utilizados em Machine Learning: K-Nearest Neighbor (KNN), Support Vector Machine (SVM), Gaussian Naïve Bayes (GNB), SOFTMAX e Random Forest (RF). Esses algoritmos possuem variáveis definidas antes do treinamento, chamadas de hiperparâmetros. As métricas do conjunto de teste foram avaliadas de acordo com: precisão, revocação (recall), resultado f1 (f1 score) e acurácia. O algoritmo classificador SoftMax apresentou um resultado de melhor comparação em relação aos demais classificadores, em todas as métricas analisadas, apresentando a matriz de confusão com melhor acurácia, com média aproximada de 35% de precisão. Deste modo, optou-se por realizar testes com uma classificação binária (hormonal e não hormonal) com o algoritmo classificador de melhor resultado no estudo, o SoftMax. Assim, atingiu-se uma porcentagem de acurácia de 65%, apresentando-se como um classificador mais promissor na classificação binária (hormonal e não hormonal) de câncer de mama. Considerações finais: A análise de ressonância magnética de câncer de mama com a utilização de uma nova Rede Neural Convolucional pode realizar uma classificação binária de câncer de mama com uma precisão de 65%. Evidências mostram que conjuntos de dados maiores levam a um maior desempenho do modelo proposto. Assim, como continuidade desta pesquisa, serão utilizados maiores números de conjuntos de dados e outros classificadores em novos testes, com intuito de obter resultados mais promissores.
Introduction: Breast Cancer is a complex and heterogeneous disease, currently classified into specific subtypes that are associated with different prognoses. Biopsy remains the gold standard for identifying biomarkers, but it has important limitations. The use of Machine Learning (ML) is not new in cancer research. Artificial neural networks (ANNs) have been used in cancer detection and diagnosis for approximately 20 years. Objectives: Thus, this research aims to develop a convolutional neural network (CNN) that can predict the molecular subtype of breast cancer based on images obtained by magnetic resonance imaging. Methods: The study was carried out at the Hospital das Clínicas da Faculdade de Medicina de Botucatu in magnetic resonance imaging of 76 female patients. The sections of magnetic resonance images were used for segmentation of tumor subtypes. An RNC architecture was designed with a customized convolutional Boltzmann machine using magnetic resonance imaging of 76 patients, containing 10 images each, which were used for training and validation of the RNC. Images were adjusted for validation. A balanced set of 40 images (examination of 4 patients) were used to define tests, with images of one patient for each tumor subtype (Luminal A, Luminal B, HER2+ and Triple Negative). The software code was written in Python on a personal workstation. Discussion and Results: Ten images of each tumor subtype were evaluated (Luminal A, Luminal B, HER2+ and triple negative). The performance of five classifier methods used in Machine Learning were evaluated: K-Nearest Neighbor (KNN), Support Vector Machine (SVM), Gaussian Naïve Bayes (GNB), SOFTMAX and Random Forest (RF). These algorithms have variables defined before training, called hyperparameters. The test set metrics were evaluated according to: precision, recall, f1 score and accuracy. The SoftMax classifier algorithm presented a result of better comparison in relation to the other classifiers, in all analyzed metrics, presenting the confusion matrix with better accuracy, with an approximate average of 35% of precision. Thus, it was decided to carry out tests with a binary classification (hormonal and non-hormonal) with the classification algorithm with the best result in the study, SoftMax. Thus, an accuracy percentage of 62% was reached, presenting itself as a more promising classifier in the binary classification (hormonal and non-hormonal) of breast cancer. Final considerations: Breast cancer magnetic resonance analysis using a new Convolutional Neural Network can perform a binary classification of breast cancer with an accuracy of 62%. Evidence shows that larger datasets lead to greater performance of the proposed model. Thus, as a continuation of this research, larger numbers of datasets and other classifiers will be used in new tests, with the aim of obtaining more promising results.
Introduction: Breast Cancer is a complex and heterogeneous disease, currently classified into specific subtypes that are associated with different prognoses. Biopsy remains the gold standard for identifying biomarkers, but it has important limitations. The use of Machine Learning (ML) is not new in cancer research. Artificial neural networks (ANNs) have been used in cancer detection and diagnosis for approximately 20 years. Objectives: Thus, this research aims to develop a convolutional neural network (CNN) that can predict the molecular subtype of breast cancer based on images obtained by magnetic resonance imaging. Methods: The study was carried out at the Hospital das Clínicas da Faculdade de Medicina de Botucatu in magnetic resonance imaging of 76 female patients. The sections of magnetic resonance images were used for segmentation of tumor subtypes. An RNC architecture was designed with a customized convolutional Boltzmann machine using magnetic resonance imaging of 76 patients, containing 10 images each, which were used for training and validation of the RNC. Images were adjusted for validation. A balanced set of 40 images (examination of 4 patients) were used to define tests, with images of one patient for each tumor subtype (Luminal A, Luminal B, HER2+ and Triple Negative). The software code was written in Python on a personal workstation. Discussion and Results: Ten images of each tumor subtype were evaluated (Luminal A, Luminal B, HER2+ and triple negative). The performance of five classifier methods used in Machine Learning were evaluated: K-Nearest Neighbor (KNN), Support Vector Machine (SVM), Gaussian Naïve Bayes (GNB), SOFTMAX and Random Forest (RF). These algorithms have variables defined before training, called hyperparameters. The test set metrics were evaluated according to: precision, recall, f1 score and accuracy. The SoftMax classifier algorithm presented a result of better comparison in relation to the other classifiers, in all analyzed metrics, presenting the confusion matrix with better accuracy, with an approximate average of 35% of precision. Thus, it was decided to carry out tests with a binary classification (hormonal and non-hormonal) with the classification algorithm with the best result in the study, SoftMax. Thus, an accuracy percentage of 62% was reached, presenting itself as a more promising classifier in the binary classification (hormonal and non-hormonal) of breast cancer. Final considerations: Breast cancer magnetic resonance analysis using a new Convolutional Neural Network can perform a binary classification of breast cancer with an accuracy of 62%. Evidence shows that larger datasets lead to greater performance of the proposed model. Thus, as a continuation of this research, larger numbers of datasets and other classifiers will be used in new tests, with the aim of obtaining more promising results.
Descrição
Palavras-chave
Machine learning, Breast neoplasms, Artificial neural networks, Artificial intelligence, Redes neurais artificiais, Aprendizado de máquina, Inteligência artificial, Câncer de mama, Carcinoma de mama