Uso de otimização binária de lobos cinza com deep learned features para classificar imagens radiográficas de covid-19

• Autor(a) principal: Lopes, Thales Ricardo de Souza
• Data de Publicação: 2023
• Tipo de documento: Trabalho de conclusão de curso
• Idioma: por
• Título da fonte: Repositório Institucional da UNESP
• Texto Completo: https://hdl.handle.net/11449/251677
• topic: COVID-19; Imagens radiográficas; Reconhecimento de padrões; ReliefF; Algoritmo binário de Lobos Cinza; Deep learned features; Radiographic images; Pattern recognition

In this work, a method based on deep learning by transfer learning is presented to perform the classification and pattern recognition in pulmonary radiographic images, representative of healthy classes and COVID-19. Thus, deep learned features from the AlexNet, Residual Neural Network, DenseNet and EfficientNet, trained on the ImageNet dataset, will be explored. The deep learned features was analyzed from different layers, such as max_pooling_3 from AlexNet, with 9216 values, avg_pool from ResNet50, with 2048 descriptors, the avg_pool of DenseNet-201, with 1920 attributes and, finally, the layer avg_pool of EfficientNet-b0, with 1280 features. The attributes was evaluated through a two-stage selection process: ranking each entry with the ReliefF algorithm and applying a threshold to reduce the number of possible combinations; application of a selection strategy wrapper, based on animal behavior, binary gray wolf optimizer, in order to find the best combinations in each subset of attributes. The discriminative power of each solution was defined by exploring ten classifiers with different heuristics. As a result, the best association occurred from the avg_pool layer of the Densenet network, SMO classifier and using only 27 attributes. This association provided an accuracy of 97.60%, an F measure of 0.976, and an AUC of 0.967. Furthermore, this solution represents a reduction of approximately 98.59% of the initial set of features which led to a higher accuracy rate when compared to the performance of the direct application of the Convolutional Neural Network with a reduced computational cost. Additionally, we believe the details presented here can contribute to the community interested in the issues explored here, supporting the development of models aimed at the diagnosis of pulmonary images of COVID-19.