An Investigation of Deep-Learned Features for Classifying Radiographic Images of COVID-19

• Autor(a) principal: Miguel, Pedro Lucas [UNESP]
• Data de Publicação: 2023
• Outros Autores: Cansian, Adriano Mauro [UNESP], Rozendo, Guilherme Botazzo [UNESP], Medalha, Giuliano Cardozo, do Nascimento, Marcelo Zanchetta, Neves, Leandro Alves [UNESP]
• Tipo de documento: Artigo de conferência
• Idioma: eng
• Título da fonte: Repositório Institucional da UNESP
• Texto Completo: http://dx.doi.org/10.5220/0012038500003467; http://hdl.handle.net/11449/248922
• topic: Convolutional Neural Networks; COVID-19; Deep-Learned Features; Radiographic Images; RelieF

In this proposal, a study based on deep-learned features via transfer learning was developed to obtain a set of features and techniques for pattern recognition in the context of COVID-19 images. The proposal was based on the ResNet-50, DenseNet-201 and EfficientNet-b0 deep-learning models. In this work, the chosen layer for analysis was the avg pool layer from each model, with 2048 features from the ResNet-50, 1920 features from the DenseNet0201 and 1280 obtained features from the EfficientNet-b0. The most relevant descriptors were defined for the classification process, applying the ReliefF algorithm and two classification strategies: individually applied classifiers and employed an ensemble of classifiers using the score-level fusion approach. Thus, the two best combinations were identified, both using the DenseNet-201 model with the same subset of features. The first combination was defined via the SMO classifier (accuracy of 98.38%) and the second via the ensemble strategy (accuracy of 97.89%). The feature subset was composed of only 210 descriptors, representing only 10% of the original set. The strategies and information presented here are relevant contributions for the specialists interested in the study and development of computer-aided diagnosis in COVID-19 images.