Modelos híbridos para classificar imagens histológicas: uma associação de deep features por transferência de aprendizado com comitê de classificadores
Carregando...
Data
2021-08-05
Autores
Oliveira, Cléber Ivo de
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Estadual Paulista (Unesp)
Resumo
O uso de modelos CNN explorando transferência de aprendizado profundo é uma estratégia que pode ser aplicada para definir atributos de alto nível e, consequentemente, permitir investigações de padrões em conjuntos reduzidos de imagens, situação comumente evidenciada no contexto de imagens médicas. A principal vantagem dessa estratégia está em projetar modelos que minimizam a ocorrência de overfitting, tornando-os mais úteis para a prática clínica. Portanto, neste trabalho, descrevemos uma proposta capaz de definir modelos híbridos para classificar imagens histológicas de tecido mamário, colorretal e hepático, por meio da associação de deep features via transferência de aprendizado, seleção por ranqueamento e classificação via comitê. Os atributos foram definidos a partir de camadas das arquiteturas AlexNet e ResNet-50. Os atributos foram organizadas em subconjuntos de características mais relevantes e avaliados por meio de validação cruzada k-fold. Os principais modelos híbridos foram definidos com deep features fornecidas pela rede ResNet-50, utilizando as camadas activation_48_relu e avg_pool. Os melhores resultados foram valores de acurácia de 98,00% e 99,32%, ao utilizar no máximo 35 deep features. O modelo foi capaz de reduzir o total de deep features em até 99,86% para obter os melhores valores de acurácia. Nossa proposta com informações detalhadas sobre os métodos, características e melhores associações são contribuições relevantes para a comunidade interessada no estudo de técnicas de machine learning para o reconhecimento de padrões.
The use of CNN models exploring deep learning transfer is a strategy that can be applied to define high-level attributes and make them feasible to carry out pattern investigations in reduced image sets, a situation commonly evidenced in the context of medical images. The obtained models can minimize the occurrence of overfitting and make them more useful for clinical practice. In this work, we describe a proposal capable of defining hybrid models to classify histological images from breast, colorectal and liver tissues, through the association of deep features by transfer learning, selection by ranking, and ensemble classifier. The features were defined by using layers from the AlexNet and ResNet-50 architectures. The attributes were organized into subsets of the most relevant features and submitted to a k-fold cross-validation process. The main hybrid models were defined with deep features from the ResNet-50 network, using the activation_48_relu and avg_pool layers. The best results were accuracy values of 98.00% and 99.32%, using a maximum of 35 deep features. The model was able to reduce the total of deep features by up to 99.86% to obtain the best accuracy values. Our proposal with detailed information regarding the methods, features and best combinations are relevant contributions for the community interested in the study of machine learning techniques for pattern recognition.
The use of CNN models exploring deep learning transfer is a strategy that can be applied to define high-level attributes and make them feasible to carry out pattern investigations in reduced image sets, a situation commonly evidenced in the context of medical images. The obtained models can minimize the occurrence of overfitting and make them more useful for clinical practice. In this work, we describe a proposal capable of defining hybrid models to classify histological images from breast, colorectal and liver tissues, through the association of deep features by transfer learning, selection by ranking, and ensemble classifier. The features were defined by using layers from the AlexNet and ResNet-50 architectures. The attributes were organized into subsets of the most relevant features and submitted to a k-fold cross-validation process. The main hybrid models were defined with deep features from the ResNet-50 network, using the activation_48_relu and avg_pool layers. The best results were accuracy values of 98.00% and 99.32%, using a maximum of 35 deep features. The model was able to reduce the total of deep features by up to 99.86% to obtain the best accuracy values. Our proposal with detailed information regarding the methods, features and best combinations are relevant contributions for the community interested in the study of machine learning techniques for pattern recognition.
Descrição
Palavras-chave
Deep features, Transferência de aprendizado, Modelos híbridos, Imagens histológicas, Reconhecimento de padrões, Transfer learning, Hybrid models, Histological images, Pattern recognition