Automated Detection of COVID-19 From CXR Image Using Voting Ensemble CNNs Transfer Learning

  1. Phuwadol Viroonluecha 1
  2. Thanwarat Borisut
  3. José Santa Lozano 1
  1. 1 Universidad Politécnica de Cartagena
    info

    Universidad Politécnica de Cartagena

    Cartagena, España

    ROR https://ror.org/02k5kx966

Journal:
Journal of Intelligent Informatics and Smart Technology

ISSN: 2586-9167

Year of publication: 2021

Type: Article

More publications in: Journal of Intelligent Informatics and Smart Technology

Abstract

Controlling the spread and investigating the COVID-19 coronavirus outbreak is a big challenge, attending to the exponentially growing number of cases. The high number of suspected cases causes medical personnel to work hard with time restrictions to perform tests. Although patients are mainly screened with Polymerase Chain reaction (PCR) tests, the chest X-ray is an effective method for detecting infections in patients, including COVID-19. In this paper, we propose to automate this virus detection by image processing using chest X-rays, using Convolutional Neural Networks (CNNs). We introduce the Voting Ensemble CNNs Transfer Learning technique, which is applied with pre-trained weights from another dataset to solve the training data insufficiency issue that can lead to low classification accuracy and strengthen a prediction result. The training and evaluation datasets were collected from several sources, including a COVID-19 image data collection and NIH Chest X-ray Dataset of 14 Common Thorax Disease Categories. Several data preparation techniques such as data augmentation and heatmap were used to improve the prediction during the experiment. After comparing 14 CNNs algorithms' performance, we chose five of the finest algorithms for the voting ensemble to build 3-voter, 4-voter and 5-voter ensembles. The best performer is the 3-voter ensemble which its voters are ResNet18, ResNet34 and AlexNet. The results show that our model outputs an accuracy of 0.9, recall of 0.825, precision of 0.971, and F1 score of 0.892. https://jiist.aiat.or.th/assets/uploads/16195381857462tOQIJIIST-40-FinalVersion.pdf