Investigação de Deep Learned Features no contexto de imagens de COVID-19

Carregando...
Imagem de Miniatura

Data

2022-12-01

Autores

Miguel, Pedro Lucas

Título da Revista

ISSN da Revista

Título de Volume

Editor

Universidade Estadual Paulista (Unesp)

Resumo

Neste trabalho, foi desenvolvido um estudo fundamentado em deep learned features por transfer learning para definir um conjunto de atributos e técnicas a fim de classificar e reconhecer padrões em imagens radiográficas de COVID-19. A análise foi realizada por meio de três modelos de aprendizado profundo. As arquiteturas escolhidas foram a Residual Neural Network com 50 camadas (ResNet-50), Dense Convolutional Network com 201 camadas (DenseNet-201) e a EfficientNet-b0 (com 237 camadas). Neste estudo, a camada avg_pool de cada arquitetura foi escolhida para a análise, com 2048 descritores na ResNet-50, 1920 na DenseNet-201 e 1280 na EfficientNet-b0. As deep learned features mais relevantes foram definidas para o processo de classificação, aplicando o algoritmo RelieF com limiares distintos e duas estratégias de classificação: classificadores aplicados individualmente e por meio de ensemble, usando a abordagem de fusão em nível de pontuação. Como parte das contribuições deste trabalho, foram identificadas duas principais combinações, ambas resultantes do modelo DenseNet-201 e um único conjunto de features mais representativo. A primeira combinação foi com o uso do classificador SMO (98,38% de acurácia) e a segunda foi do ensemble (acurácia de 97,89%). O conjunto de atributos mais relevante envolveu somente 210 deep learned features, com apenas 10% do total inicial de atributos. As informações expostas são contribuições relevantes para os especialistas interessados no estudo e desenvolvimento de sistemas direcionadas para imagens radiográficas representativas de COVID-19.
In this proposal, a study based on deep learned features from transfer learning was conducted to obtain a set of features and techniques to classify and recognize patterns in COVID-19 images. The analyses was conducted using a total of three deep learning models. The chosen models were Residual Neural Network using 50 layers (ResNet50), Dense Convolutional Network using 201 layers (DenseNet-201) and EfficientNetb0 using 237 layers. In this proposal, the chosen layer for analysis was the avg_pool layer from each model, this layer is composed by 2048 features on ResNet-50, 1920 features on DenseNet0201, and finally, 1280 features on the EfficientNet-b0. The most relevant deep learned features where obtained for the classification process, applying the RelieF algorithm and two classification strategies: classifiers applied individually and through ensemble of classifiers using score level fusion. As part of the contributions obtained from this study, two combinations were identified, both of them using the DenseNet-201 model with the same set of features. The first combination used the SMO classifier (accuracy of 98,38%) and the second combination used ensemble (accuracy of 97,89%). The most relevant set of features was composed with only 210 deep learned features, using only 10% of the initial amount of features. Those informations are relevant contributions to specialists that are interested in the study and development of systems oriented to X-ray images of COVID-19.

Descrição

Palavras-chave

COVID-19 (Doença), Redes neurais (Computação), Inteligência artificial Aplicações médicas, Processamento de imagens, COVID-19 (Disease), Neural networks (Computer science), Artificial intelligence Medical applications, Image processing

Como citar