Exploração de transformada Curvelet com redes neurais convolucionais para classificação de imagens histológicas

Carregando...
Imagem de Miniatura

Data

2021-12-06

Autores

Gervaes, Guilherme

Título da Revista

ISSN da Revista

Título de Volume

Editor

Universidade Estadual Paulista (Unesp)

Resumo

Feature extraction and the method utilized for data processing are fundamental parts of the digital image classification process. Therefore, in this work, the classification capacity of a convolutional neural network was observed when curvelet coefficients extracted from histological images were added to the original histological images, in contexts of colorectal cancer, non-Hodgkin lymphomas, breast cancer and liver tissue of mice of both genders and at different stages of life in order to observe the best combinations of histological images and curvelet coefficients in the classification process. In the feature extraction step, curvelet coefficients of histological images were obtained and sub-images composed of these coefficients were generated. The classification, performed by the pre-trained residual convolutional neural network ResNet50, used the images obtained in the feature extraction step combined with the classification using only histological images. The most promising results were observed when the sub-images generated from the coefficients were added to the original histological images, producing results up to 10.70% superior in accuracy and 25% superior in precision when compared to those observed in the classification using only histological images. The combination of the original histological images with the sub-images generated from the curvelet coefficients presents a relevant contribution for researchers interested in the development of computer-aided systems that help professionals in decision-making in the domains presented in the work.
A extração de características e o método utilizado para processamento dos dados são partes fundamentais do processo de classificação de imagens digitais. Portanto, neste trabalho, foi observada a capacidade de classificação de uma rede neural convolucional quando adicionados coeficientes curvelet extraídos de imagens histológicas às imagens histológicas originais, em contextos de câncer colorretal, linfomas Não-Hodgkin, câncer de mama e tecido do fígado de ratos de ambos os gêneros e em diferentes estágios de vida, com objetivo de observar as melhores combinações de imagens histológicas e coeficientes curvelet no processo de classificação. Na etapa de extração de características foram obtidos os coeficientes curvelet das imagens histológicas e geradas sub-imagens compostas por esses coeficientes. A classificação, realizada pelo modelo pré-treinado da rede neural convolucional residual ResNet50, utilizou as imagens obtidas na extração de características combinadas às imagens histológicas e o desempenho foi comparado com a classificação utilizando apenas imagens histológicas. Os resultados mais promissores foram observados quando as sub-imagens geradas a partir dos coeficientes foram adicionadas às imagens histológicas originais, produzindo resultados até 10,70% superiores em acurácia e 25% superiores em precisão quando comparados aos observados na classificação utilizando somente imagens histológicas. A combinação das imagens histológicas originais com as sub-imagens geradas a partir dos coeficientes curvelet apresenta contribuição relevante para pesquisadores que se interessam pelo desenvolvimento de sistemas de apoio a diagnóstico que auxiliem profissionais na tomada de decisão nos domínios apresentados no trabalho.

Descrição

Palavras-chave

Feature extraction, Curvelet transform, Extração de características, Transformada curvelet, ResNet50

Como citar