Proposing a Deep Learning Based Solution for Detecting Suspicious Cases of COVID-19
Subject Areas : مهندسی برق و کامپیوتر
Atena Abidi
1
,
Haniye Jamahmoodi
2
,
Zahra Heydaran Daroogheh Amnyieh
3
,
iman zabbah
4
1 - Dept. of Comp. Eng., Bushehr Branch , Islamic Azad University, Bushehr, Iran
2 - Dept. of Comp. Eng., Mashhad Branch , Islamic Azad University, Mashhad, Iran
3 - Dept. of Elec. Eng., Dolatabad Branch, Islamic Azad University, Isfahan, Iran
4 - Dept. of Comp. Eng., Torbat Heydariyeh Branch, Islamic Azad University, Torbat Heydariyeh, Iran
Keywords: COVID-19, Deep learning, Data mining,
Abstract :
Deep neural networks are used in the detection of diseases and medical tasks due to their power and capability in extracting complex features and non-linear relationships. Following the emergence of COVID-19, deep learning approaches have been introduced as a powerful approach in diagnosing this disease. In some cases, data mining-based methods cannot definitively diagnose COVID-19 due to their lack of appropriate generalizability on the data. The aim of this research is to propose a solution to improve the diagnostic results in suspicious COVID-19 images.
In this study, after diagnosing the disease using two deep networks, GoogleNet and AlexNet, the probability layer of the two learned networks is extracted, and the suspicious cases of the disease are identified. Then, the top features extracted from the two deep learners are combined and sent to a perceptron neural network for the diagnosis of suspicious cases. The extraction of the best features was performed using principal component analysis. The study database includes 224 CT scan images of COVID-19-infected lungs and 522 lung images of healthy individuals, obtained from the GitHub repository. The study results indicate that the aggregation of deep learners in the probability layer can lead to a 98.1% improvement in the diagnosis of COVID-19 in suspicious cases.