Detection of Covid-19 through the Analysis of Radiographic Images of the Chest Using Convolutional Neural Networks

Authors

Jácome Morales, Gladys

Cedeño Rodríguez, Juan

Patiño-Pérez, Darwin

Collantes Farah, Alex

Burgos Robalino, Freddy

Pazmiño Moran, Victor

Molina Calderón, Miguel

Abstract

There are different ways to detect Covid-19, which have emerged so far giving an effective response in detecting the disease, the PCR test is a reliable diagnostic method, which requires a well-equipped laboratory to obtain results, which can take hours or days. Another detection technique for this disease is by analyzing the chest image; This technique is used as a diagnostic tool in emergency areas in health centers, because it can reveal characteristics related to lung involvement. For this reason, it is important to develop an automatic detection system, as an alternative diagnosis option for Covid-19. Deep Learning techniques can help detect the SARS-CoV-2 virus by analyzing chest radiographic images. Thanks to the high availability of the datasets available, and using convolutional neural networks, the analysis is carried out by classifying images. In this research, two CNN models were created whose outputs are normal or covid19, the same ones that were trained with two datasets from public research repositories. The performance of the models trained in Pytorch were compared with the models trained in Keras under similar conditions of parameters and hyperparameters, obtaining a higher performance with Pytorch however since the two types of models have learned adequately with an accuracy that is above the 90% recommended the use of both models.