Amirkabir University of Technology AUT Journal of Electrical Engineering 2588-2910 53 2 2021 12 01 A Hybrid Deep Transfer Learning-based Approach for COVID-19 Classification in Chest X-ray Images 223 232 4305 10.22060/eej.2021.19467.5397 EN Khosro Rezaee Department of Biomedical Engineering, Meybod University, Meybod, Iran 0000-0001-6763-6626 Afsoon Badiei Department of Biomedical Engineering, Faculty of Electrical and Computer Engineering, Tabriz University, Tabriz, Iran Hossein Ghayoumi Zadeh Department of Electrical Engineering, Faculty of Engineering, Vali-e-Asr University of Rafsanjan, Rafsanjan, Iran 0000-0002-5390-3938 Saeed Meshgini Department of Biomedical Engineering, Faculty of Electrical and Computer Engineering, Tabriz University, Tabriz, Iran 0000-0001-5023-0961 Journal Article 2021 01 08 The COVID-19 pandemic is a severe public health hazard. Hence, proper and early diagnosis is necessary to control the infection progression. We can diagnose this disease by employing a chest X-ray (CXR) screening, which is ordinarily cheaper and less harmful than a Computed Tomography scan (CT scan) and is continuously accessible in small or rustic hospitals. Since the COVID-19 dataset is inadequate and cannot be strictly distinguished from CXR, Deep Transfer Learning (DTL) models can be used to diagnose coronavirus even with access to a small number of images. In this paper, we presented an approach to diagnosis COVID-19 using CXR images based on the concatenated features vector of the three DTL structures and soft-voting feature selection procedure, including Receiver of Curve (ROC), Entropy, and signal-to-noise ratio (SNR) techniques. Our hybrid model reduces the feature vector size and classifies it in optimize manner to improve the decision-making process. A collection of 2,863 CXR images comprising normal, bacterial, viral, and COVID-19 cases were prepared in JPEG format from the Medical Imaging Center of Vasei Hospital, Sabzevar, Iran. The proposed approach obtained an Accuracy of 99.34%, Sensitivity of 99.48%, Specificity of 99.27% while having a far fewer number of trainable parameters in contrast to its counterparts. Compared to the latest similar methods, the diagnosis accuracy has increased from 1.5 to 2.2%. The comparative experiment reveals the advantage of the suggested COVID-19 classification pattern based on DTL over other competing schemes. https://eej.aut.ac.ir/article_4305_d6cf4da5ced8580c991e16fb54faa1b6.pdf