Gustavo Chichanoski and Maria Bernadete de Morais França, State University of Londrina, Brazil

Covid-19 is caused by the SARS-COV2 virus, where most people experience a mild to moderate respiratory crisis. To assist in diagnosing and triaging patients, this work developed a Covid-19 classification system through chest radiology images. For this purpose, the neural network models ResNet50V2, ResNet101V2, DenseNet121, DenseNet169, DenseNet201, InceptionResnetV2, VGG-16, and VGG-19 were used, comparing their precision, accuracy, recall, and specificity. For this, the images were segmented by a U-Net network, and packets of the lung image were generated, which served as input for the different classification models. Finally, the probabilistic Grad-CAM was generated to assist in the interpretation of the results of the neural networks. The segmentation obtained a Jaccard similarity of 94.30%, while for the classification the parameters of precision, specificity, accuracy, and revocation were evaluated, compared with the reference literature. Where DenseNet121 obtained an accuracy of 99.28%, while ResNet50V2 presented a specificity of 99.72%, both for Covid-19.

Deep Learning, U-Net, Covid-19, Segmentation & Machine Learning.