Arquiteturas de Redes Neurais Convolucionais para Identificação de Pneumonia e COVID-19 em Raios X de Tórax
Keywords: Convolutional Neural Networks, X-Ray, COVID-19, Pneumonia
AbstractThis work compares architectures of convolutional neural networks for the identification of pneumonia and COVID-19 in chest X-rays. The models considered in the evaluation were ResNet-50, MobileNet-v2, Inception-v3, and EfficientNet-B2, besides ensembles built using these neural networks. We also propose and evaluate a preprocessing technique to reduce variations in the images. Experimental results show that MobileNet-v2 achieves the highest individual performance with accuracy of 94.03%, precision of 94.59%, recall of 91.91%, and f1-score of 91.55%. The ensemble with highest performance achieved accuracy of 96.00%, precision of 94.61%, recall of 95.53%, and f1-score of 95.04%. Integrated gradients analyses showed that models focus on regions that make sense when performing the predictions.