COVID-19 Detection on Chest X-Ray Images Using Modified VGG-19

Abstract

A global pandemic infecting human respiratory system identified as COVID-19 emerged in 2019 has been a critical issue concerning worldwide, including Indonesia. The massive and relatively easy spread of the virus requires COVID-19 patients to be separated from other patients to avoid hospitals being the major source of the virus spread. The gold standard in COVID-19 testing known as rRt-PCR was established to help detecting the virus. This method, however, is found to be less time-efficient due to the sample examination taking 5-7 hours to conduct. Another COVID-19 testing method called Rapid Antigen Test is claimed to be more practical since the sample examination only takes 5-30 minutes to complete. One disadvantage of this testing method is its sensitivity, which is believed to be relatively low. Further supporting examinations are needed in diagnosing COVID-19 when using the Rapid Antigen Test, one of which is chest imaging. Most hospitals in Indonesia are type C hospitals, which means radiology specialists’ availability is limited and sometimes not available. Therefore, it is necessary to develop an automation system for detecting COVID-19 disease on chest X-ray images using the Convolutional Neural Network method to help doctors interpreting chest X-rays of patients effectively faster in determining whether the patient is potentially infected with COVID-19 or not. The algorithm proposed in this study was modified VGG-19, with ADAM optimization method and categorical cross-entropy loss function. This algorithm resulted in a model accuracy rate of 81%, a precision/recall model of 62%, and a specificity model of 87%. The area under the curve in the Receiver Operatic Characteristic Curve indicates that the accuracy level of the test is below the Excellent curve, suggesting that the model created successfully serves as a screening test tool.