Ensemble learning como estratégia para investigar imagens H&E utilizando duplo estágio de seleção de atributos
Carregando...
Data
2023-02-01
Autores
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Estadual Paulista (Unesp)
Resumo
Neste trabalho, é apresentada uma investigação de ensemble learning para o contexto de imagens médicas, especificamente o reconhecimento de padrões em amostras histológicas tingidas com Hematoxilina e Eosina. As bases exploradas foram representativas do câncer colorretal, displasia oral epitelial, linfomas não-Hodgkin e tecidos hepáticos. A estratégia de ensemble learning foi considerada a partir de múltiplos descritores, tais como deep learned e handcrafted, e múltiplos classificadores. Os descritores deep learned foram calculados explorando cinco distintas arquiteturas de redes neurais convolucionais. Os descritores handcrafted foram representativos das categorias fractais multidimensionais e multiescala, Haralick e Local Binary Pattern. As principais combinações de descritores, consequentemente de técnicas, foram obtidas por meio de duplo estágio de seleção de atributos (ranqueamento com meta-heurísticas) e classificadas via um ensemble de classificadores compostos por múltiplos classificadores: Support Vector Machine, Naive Bayes, Random Forest, Regressão Logística e K-Nearest Neighbors. As taxas de acurácias foram valores entre 90,72% e 100,00%, com alguns destaques envolvendo combinações de técnicas capazes de aprimorar as acurácias em diferentes contextos de imagens histológicas; análise dos atributos presentes nas melhores soluções obtidas; e definição de combinações de ensembles com desempenhos competitivos em relação aos disponíveis na literatura, explorando um número reduzido de atributos.
In this work, an ensemble learning investigation is presented for the context of medical images, specifically the recognition of patterns in histological samples stained with Hematoxylin and Eosin. The datasets explored were representative of colorectal cancer, oral epithelial dysplasia, non-Hodgkin lymphoma, and liver tissue. The ensemble learning strategy was considered from multiple descriptors, such as deep learned and handcrafted, and multiple classifiers. The deep learned features were calculated by exploring five different architectures of convolutional neural network. The handcrafted features were representative of the multidimensional and multiscale fractal, Haralick and Local Binary Pattern categories. The main combinations of descriptors, consequently of techniques, were obtained through a two-stage feature selection (ranking with metaheuristics) and classified via an ensemble of classifiers composed of multiple classifiers: Support Vector Machine, Naive Bayes, Random Forest, Logistic Regression, and K-Nearest Neighbors. The accuracy rates were values between 90.72% and 100.00%, with some highlights involving combinations of techniques capable of improving accuracy in different histological image contexts; analysis of the features present in the best solutions obtained; and definition of ensemble combinations with competitive performances compared to those available in the literature, exploring a reduced number of descriptors.
In this work, an ensemble learning investigation is presented for the context of medical images, specifically the recognition of patterns in histological samples stained with Hematoxylin and Eosin. The datasets explored were representative of colorectal cancer, oral epithelial dysplasia, non-Hodgkin lymphoma, and liver tissue. The ensemble learning strategy was considered from multiple descriptors, such as deep learned and handcrafted, and multiple classifiers. The deep learned features were calculated by exploring five different architectures of convolutional neural network. The handcrafted features were representative of the multidimensional and multiscale fractal, Haralick and Local Binary Pattern categories. The main combinations of descriptors, consequently of techniques, were obtained through a two-stage feature selection (ranking with metaheuristics) and classified via an ensemble of classifiers composed of multiple classifiers: Support Vector Machine, Naive Bayes, Random Forest, Logistic Regression, and K-Nearest Neighbors. The accuracy rates were values between 90.72% and 100.00%, with some highlights involving combinations of techniques capable of improving accuracy in different histological image contexts; analysis of the features present in the best solutions obtained; and definition of ensemble combinations with competitive performances compared to those available in the literature, exploring a reduced number of descriptors.
Descrição
Palavras-chave
Ensemble learning, Descritores handcrafted, Descritores deep learned, Imagens histológicas, Duplo estágio de seleção de atributos, Handcrafted features, Deep learned features, Histological images, Two-stage feature selection