Open AccessFully-Connected Neural Network for COVID-19 Chest X-Ray Imaging Classification Using Hybrid Features
Author(s) -
Victor Hugo Viveiros,
Rayanne Lima,
Fernando Lucas Martins,
Alessandra Martins Coelho,
Matheus de Freitas Oliveira Baffa
Publication year - 2020
Language(s) - English
Resource type - Conference proceedings
DOI - 10.5753/wvc.2020.13498
Subject(s) - computer science , pneumonia , artificial intelligence , feature extraction , artificial neural network , medical imaging , covid-19 , transmission (telecommunications) , pandemic , computer vision , pattern recognition (psychology) , medicine , pathology , disease , telecommunications , infectious disease (medical specialty)
Discovered on 31st December of 2019, the new Coronavirus has a high transmission capacity and was considered pandemic by the World Health Organization. In only six months is was able to spread all over the world and cause more than 600 thousand deaths. Early diagnosis is essential for governments to take public policies, such as social isolation, commerce control, and contact tracking. In order to make these actions, massive tests are required. On the other hand, diagnosis kits are expensive and not accessible to everyone. Medical imaging, such as thoracic x-ray and Computational Tomography (CT) has been used to visualize the lung and to verify at the first moment the presence of viral pneumonia. However, some countries have few radiologists specializing in chest x-ray analysis. The findings in the image are generally not so easy to see and can easily be confused with traditional pneumonia findings. For this reason, studies in Computer Vision are necessary, both to detect anomalies in imaging and to differentiate the other types of pneumonia. This paper addresses the initial results of a research, which developed an image classification methodology to differentiate x-ray images from sick patients, infected with Coronavirus, and healthy patients. The proposed method, based on the extraction and detection of patterns in texture and color features through a Deep Neural Network, obtained an average accuracy of 95% following a k-fold cross-validation experiment.