Exploração de deep features via transfer learning com técnicas fractais para classificar imagens H&E
Carregando...
Data
2022-12-12
Autores
Longo, Leonardo Henrique da Costa
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Estadual Paulista (Unesp)
Resumo
Neste trabalho é apresentada uma proposta para investigar combinações de descritores handcrafted e deep learned, bem como possíveis padrões de associações em diferentes tipos de imagens histológicas. Os descritores handcrafted foram definidos da categoria de técnicas fractais aplicadas às imagens originais, bem como às representações obtidas por meio de técnicas de inteligência artificial explicável. Os descritores deep learned foram obtidos de diferentes arquiteturas de redes neurais convolucionais. Os descritores mais relevantes de cada combinação, definidos a partir de um algoritmo de ranqueamento, foram analisados em um ensemble composto pelos classificadores SVM, Naive Bayes, Random Forest e K-Nearest Neighbors. A metodologia proposta foi aplicada em imagens H&E representativas do câncer de mama, colorretal, displasia oral e tecido hepático. Os melhores resultados em cada conjunto foram taxas de acurácias de 94,83% a 100%, além de: conhecer padrões de combinações de técnicas para diferentes tipos de imagens H&E; indicar que o ensemble de descritores permite aumentar as taxas conquistadas em diferentes contextos, sobretudo quando combinadas com deep features obtidas via transfer learning; definir que combinação de ensembles (descritores e classificadores) pode fornecer desempenhos competitivos em relação aos disponíveis na Literatura, explorando um número reduzido de características.
In this project it is proposed a methodology to investigate the ensemble of handcrafted and deep learned features, as well as to find possible patterns in multiple associations on images from different histological origins. The handcrafted features were selected from fractal techniques, applied to the original images as well to their respective explanations obtained with explainable artificial inteligence approaches. On the other hand, the deep learned features were extracted from several convolutional neural networks. The most relevant features from each ensemble, according to a ranking algorithm, were fed to and ensemble composed by the classifiers SVM, Naive Bayes, Random Forest and K-Nearest Neighbors. Our proposed method was performed with H&E images representing breast cancer, colorrectal cancer, oral displasia and liver tissue. The best results from each context achieved accuracies ranging from 94.83% to 100%, besides that we also: found a pattern for ensemble techniques considering multiple types of H&E images; demonstrated that the ensemble of features allows to improve accuracy on several scenarios, specially with deep learned features acquired with transfer learning; defined which ensemble combination (features and classifiers) could provide competitive results agains other proposed methods on the literatures while using a small number os features.
In this project it is proposed a methodology to investigate the ensemble of handcrafted and deep learned features, as well as to find possible patterns in multiple associations on images from different histological origins. The handcrafted features were selected from fractal techniques, applied to the original images as well to their respective explanations obtained with explainable artificial inteligence approaches. On the other hand, the deep learned features were extracted from several convolutional neural networks. The most relevant features from each ensemble, according to a ranking algorithm, were fed to and ensemble composed by the classifiers SVM, Naive Bayes, Random Forest and K-Nearest Neighbors. Our proposed method was performed with H&E images representing breast cancer, colorrectal cancer, oral displasia and liver tissue. The best results from each context achieved accuracies ranging from 94.83% to 100%, besides that we also: found a pattern for ensemble techniques considering multiple types of H&E images; demonstrated that the ensemble of features allows to improve accuracy on several scenarios, specially with deep learned features acquired with transfer learning; defined which ensemble combination (features and classifiers) could provide competitive results agains other proposed methods on the literatures while using a small number os features.
Descrição
Palavras-chave
Processamento de imagens, Aprendizado do computador, Redes neurais (Computação), Fractais, Deep features, Fractal techniques, Feature ensemble, Ensemble of classifiers, H&E histological images