Abstract
The world has experienced the very first pandemic of 21st century, called the COVID-19 which is caused by a deadly virus named Coronavirus. In this regard, one of the very first strategy to minimize the number of affected patients and reduce casualties is to diagnose COVID-19 at an early stage. Currently, PCR test is primarily utilized for the diagnosis of COVID-19. However, PCR test requires a huge number of expensive test kits as well as trained experts. Therefore, chest X-Ray imaging technique (including Machine Learning) has been considered as an alternative for COVID-19 diagnosis among the researchers. This particular method is faster, less expensive and will allow the authorities to manage the COVID19 diagnosis system in a cost-effective way. Machine learning techniques have been proven to be significantly efficient and accurate for image classification problems. On the other hand, One of the most utilized techniques in machine learning is supervised learning which is highly convenient and helping the experts to diagnose and make informed decisions about COVID-19. Supervised learning in image classification requires vast amount of radiography images with notable accuracy which can be a peculiar issue in medical domain. In order to address the problem, we have investigated a distinct approach for COVID-19 Diagnosis with a nominal dataset. In this work, We have studied the effectiveness of Semi-Supervised Learning (SSL) for COVID-19 diagnosis from chest X-ray images. We have investigated a prepossessing technique by extracting and combining local phase image feature into multi-feature image to train our SSL model in teacher/student archetype. Our study have shown that by using 17.0% of the total dataset for training, the SSL model achieve 93.45% accuracy. We also provide comparative metrics of SSL approach against other fully supervised techniques.