Resumen

LOPEZ-BETANCUR, Daniela et al. Comparison of Convolutional Neural Network Architectures for COVID-19 Diagnosis. Comp. y Sist. [online]. 2021, vol.25, n.3, pp.601-615.  Epub 13-Dic-2021. ISSN 2007-9737.  https://doi.org/10.13053/cys-25-3-3453.

Convolutional neural networks (CNNs) have shown great potential to solve several medical image classification problems. In this research, thirty-two CNN architectures were evaluated and compared to perform COVID-19 diagnosis by using radiographic images. A collection of 5,953 frontal chest X-ray images (117 patients diagnosed with COVID-19, 4,273 with Pneumonia not related to COVID-19, and 1,563 Normal or healthy) was used for training and testing those thirty-two architectures. In this article, the implemented metrics were according to the conditions of an imbalanced dataset. Seven of the thirty-two models evaluated achieved an excellent performance classification (≥90%) according to the Index of Balanced Accuracy (IBA) metric. The top three CNN models selected in this research (Wide_resnet101_2, Resnext101_32x8d, and Resnext50_32x4d) obtained the highest classification precision value of 97.75%. The overfitting problem was ruled out according to the evolution of the training and testing precision measurement. The best CNN model for COVID-19 diagnosis is the Resnext101_32x8d according to the confusion matrix and the metrics achieved (sensitivity, specificity, F1-score, G_mean, IBA, and training time of 97.75%, 96.40%, 97.75%, 97.06%, 94.34%, 76.98 min, respectively) by the CNN model.

Palabras llave : Convolutional neural network; COVID-19; Transfer learning.