2021 3rd East Indonesia Conference on Computer and Information Technology (EIConCIT) April 09-11. 2021, ISTTS Surabaya, Indonesia 978-1-6654-0514-0/21/$31.00 ©2021 IEEE 343 Improvement of Xception-ResNet50V2 Concatenation for COVID-19 Detection on Chest X-Ray Images 1 st Hajar Indah Fitriasari Department of Electrical Engineering Faculty of Engineering Universitas Indonesia Depok, Indonesia hajar.indah@ui.ac.id 2 nd Mia Rizkinia Department of Electrical Engineering Faculty of Engineering Universitas Indonesia Depok, Indonesia mia@ui.ac.id Abstract—Novel coronavirus disease (COVID-19) has globally become a pandemic since the first quarter of 2020 whose growth rate must immediately be controlled. One of strategy to reduce the growth rate of sufferers is to break the chain of the spread by detecting it and carrying out quarantine. X-ray imaging can be used as a modality to detect the COVID- 19 in suspected patients’ lungs as a clinical diagnostic tool. One of the challenges of this task is the difficulty in distinguishing the characteristics of COVID-19 from other diseases with similar features of the images resulted from the X-ray of the chest. To reduce the problems that will be faced, machine learning or deep learning is embedded in an automatic computer-aided diagnosis (CAD) to increase efficiency and accuracy. Several deep learning-based artificial intelligence systems can be used in diagnosis, one of the most popular is using the previously proposed Convolutional Neural Network (CNN), which has promising accuracy in detecting COVID-19 confirmed patients using CXR images. In this study, to detect COVID-19 confirmed patients by classifying them into 4 classes, we propose a modified combination of two CNN architectures named Xception and ResNet50V2 which makes the system powerful using multiple feature extraction capabilities. The proposed method achieves high accuracy, precision, recall, and F1-Score, reaching 93.412%, 96.6%, 99.6%, and 98%, respectively. Overall, the proposed method can be used as an automatic diagnosis system that can be utilized by clinical practitioners and radiologists to diagnose, validate, and follow-up of COVID-19 suspected cases. Keywords—Deep Learning, Convolutional Neural Network (CNN), CXR Images, COVID-19. I. INTRODUCTION At the end of December 2019, there was a pneumonia cluster with an unknown etiology that was first identified epidemiologically that the virus is very similar to the pneumonia virus [1]. From samples of the lower respiratory tract, indicating that there is a new type of coronavirus, which is named Severe Acute Respiratory Coronavirus-2 (SARS- CoV-2) on February 11, 2020 announced by World Health Organization [2]. Several strategies were conducted to reduce the very fast growth rate of sufferers, such as conducting cluster investigations by tracking people who may have contact with COVID-19 patients [3], carrying out regional quarantine, and finally carrying out randomized tests to the general public with a certain number every day. Currently, several tests have been used for early detection in COVID-19 suspected cases such as using one of the types of nucleic acid amplification tests (NAAT), real-time reverse-transcription polymerase chain reaction (rRT-PCR) [4], a serological test available to detect patient’s antibodies [5], detection of antigen which detecting the presence of the virus itself by seeking at the human immune response to infection [6]. However, some of early detection methods can only be used to determine whether a test specimen is confirmed for COVID-19 or not. Further action is needed to see the severity of confirmed COVID-19 patients, one of which is manual interpretation of Chest X-ray (CXR) images by professional radiologists. The challenge faced when interpreting it manually is to be able to carefully distinguish the characteristics of COVID-19 infection with other diseases. However, this method is ineffective due to the growth rate of the number of confirmed patients which makes it possible for radiologists to be unable to complete the diagnosis on time [7]. To reduce the problems that will be faced by radiologists, machine learning or deep learning that is embedded in an automatic computer-aided diagnosis (CAD) can be used to increase efficiency and accuracy. Several deep learning models such as recurrent neural networks (RNN), deep convolutional neural networks (DNN), recursive neural networks, transfer learning, and so on that have been applied to seek features of the disease using radiological images automatically [8]. Previously, several deep learning-based artificial intelligence systems can be used in the detection, especially using the Convolutional Neural Network (CNN) that has been proposed, which had promising accuracy in detecting COVID-19 confirmed patients using CXR images. The successful use of deep learning is not based on manual feature extraction but this algorithm automatically performs feature learning from the data itself [9]. There have been several deep learning approaches, especially the use of a CNN related to the interpretation of CXR images to detect COVID-19 infection. First, there is COVID-Net, a deep CNN model to detect COVID-19 cases from CXR images sourced with five open-source databases proposed by Wang and Wong [10]. This model introduces a light weight projection – expansion – projection - extension (PEPX) to increase the representation capacity and reduce the complexity of the computations performed [10]. However, the proposed method or design is not ready for production and further testing with new data is available in the open-source database. Khan et al [11] was proposed a deep learning design to detect COVID-19 cases from CXR images using CNN utilizes one of ImageNet's pre-trained architectural models, Xception, which is tested to perform multiclass classification 2021 3rd East Indonesia Conference on Computer and Information Technology (EIConCIT) | 978-1-6654-0514-0/20/$31.00 ©2021 IEEE | DOI: 10.1109/EIConCIT50028.2021.9431916