Sistema de visão computacional para auxílio ao diagnóstico de COVID-19

Abstract

This work aims to perform a comparative analysis of COVID-19 image classification techniquesusing three pre-trained neural network architectures (Inception V3, VGG16, and ResNet50) forfeature extraction and three classifiers (KNN, MLP, and SVM) for classification. The datasetused was obtained from Kaggle and consists of chest X-ray images classified into three classes:normal, pneumonia, and COVID-19.Initially, the pre-trained neural network architectures were used to extract features from theimages, leveraging the pre-trained weights from the ImageNet database. The extracted featureswere then fed into the KNN, MLP, and SVM classifiers. Various evaluation metrics, includingconfusion matrix, precision, accuracy, F1-score, and recall, were calculated to analyze theperformance of the classifiers.Furthermore, an ensemble approach was employed by combining the two best performingclassifiers, which were SVM and MLP. However, it was observed that MLP outperformed theensemble of both classifiers. This finding suggests that, for this specific dataset, MLP was ableto capture the relevant information more efficiently.The obtained results demonstrated the effectiveness of pre-trained neural network architecturesfor feature extraction in COVID-19 images, as well as the classifiers’ ability to correctly classifythe images. Moreover, the ensemble approach showed promise in improving classificationperformance, although the results indicated that MLP alone achieved even better performance.This study contributes to the advancement of COVID-19 detection from chest X-ray images,presenting a comparative analysis of different neural network architectures, classifiers, andensemble techniques. The utilized evaluation metrics provide a comprehensive understanding ofthe model’s performance, enabling the selection of the most suitable approach for COVID-19image classification.