Entropias fuzzy e multidimensionais para quantificação de imagens histológicas colorretais
Carregando...
Data
2018-05-14
Autores
Santos, Luiz Fernando Segato dos
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Estadual Paulista (Unesp)
Resumo
Neste trabalho são apresentados dois novos métodos para calcular as entropias de imagens coloridas. Os métodos foram baseados nas entropias Amostral e de Shannon com associações de abordagens multiescala e multidimensional. Os modelos foram aplicados para quantificar e classificar imagens histológicas H&E de câncer colorretal. A abordagem multiescala foi definida com observações em janelas de diferentes tamanhos e variações da tolerância para determinar a similaridade de padrões. A estratégia multidimensional foi definida considerando cada pixel da imagem colorida no padrão RGB como um vetor n-dimensional, o qual foi analisado a partir da distância de Minkowski. A entropia Amostral também contemplou uma estratégia Fuzzy, uma função gaussiana para verificar as pertinências das distâncias das janelas. Os resultados foram métodos capazes de computar as similaridades entre pixels contidos em janelas de diversos tamanhos, bem como as informações presentes nos canais de cores. O poder de quantificação de cada modelo foi testado em uma base pública de imagens colorretais H&E, composta pelas classes benigna e maligna. Os resultados foram dados como entradas para classificadores de diferentes categorias e analisados via validação cruzada k-folds, com k=10. Os desempenhos indicaram que a associação proposta foi capaz de distinguir os grupos benigno e maligno, com valores que superaram os resultados obtidos com importantes técnicas disponíveis na literatura. O melhor desempenho foi uma AUC de 0,952, resultado que é relevante principalmente quando consideramos as dificuldades da prática clínica para separar os grupos estudados.
In this study, two new methods are presented to calculate the entropies of colour images. The methods was based on the combination of Sample and Shannon Entropy with multiscale and multidimensional approaches. The model was applied to quantify and classify histological images H&E of colorectal cancer. The multiscale approach wasdefinedbyanalysingwindowsofdifferentsizesandvariationsintolerancefordetermining pattern similarity. The multidimensional strategy was established through a consideration of each pixel in the colour image through the RGB standard, as an ndimensional vector, which was analysed from the Minkowski distance. A Fuzzy strategywasalsoconsideredbytheSampleEntropy,aGaussianfunctionusedtoverifythe pertinence of the distances between windows. The results were methods capable of computing similarities between pixels contained in windows of various sizes, as well as the information present in the colour channels. The power of quantification of each model was tested in a public colorectal H&E image database, which was composed of both benign and malignant classes. The results were given as inputs for classifiers of different categories and analysed by applying k-fold cross-validation, with k=10. The derived performances indicate that the proposed association was capable of distinguishing the benign and malignant groups, with values that surpassed those results obtained with important techniques available in the Literature. The best performance was an AUC of 0.952, a result considered relevant, mainly when we consider the difficulties of clinical practice, when it comes to separating the groups under study.
In this study, two new methods are presented to calculate the entropies of colour images. The methods was based on the combination of Sample and Shannon Entropy with multiscale and multidimensional approaches. The model was applied to quantify and classify histological images H&E of colorectal cancer. The multiscale approach wasdefinedbyanalysingwindowsofdifferentsizesandvariationsintolerancefordetermining pattern similarity. The multidimensional strategy was established through a consideration of each pixel in the colour image through the RGB standard, as an ndimensional vector, which was analysed from the Minkowski distance. A Fuzzy strategywasalsoconsideredbytheSampleEntropy,aGaussianfunctionusedtoverifythe pertinence of the distances between windows. The results were methods capable of computing similarities between pixels contained in windows of various sizes, as well as the information present in the colour channels. The power of quantification of each model was tested in a public colorectal H&E image database, which was composed of both benign and malignant classes. The results were given as inputs for classifiers of different categories and analysed by applying k-fold cross-validation, with k=10. The derived performances indicate that the proposed association was capable of distinguishing the benign and malignant groups, with values that surpassed those results obtained with important techniques available in the Literature. The best performance was an AUC of 0.952, a result considered relevant, mainly when we consider the difficulties of clinical practice, when it comes to separating the groups under study.
Descrição
Palavras-chave
Entropia amostral, Entropia de Shannon, Abordagem multiescala, Abordagem multidimensional, Função Fuzzy, Câncer colorretal, Sample entropy, Shannon entropy, Multiscale approach, Multidimensional approach, Fuzzy functions, Colorectal cancer