Pneumonia X-ray Imaging Classification Based on an Interpretable Machine Learning Model

The outbreak of Covld-19 has put tremendous pressure on medical systems around the world. The highly infectious nature of this respiratory disease challenges advanced diagnostic technology to achieve rapid, scalable, affordable, and high-precision testing. In previous studies, Tsiknakis used Convolutional Neural Network (CNN) and transfer learning to achieved high accuracy in distinguishing the lung X-ray images of Covid-19 infectors and healthy people. However, its accuracy is not so high in quaternary classification (Bacterial Pneumonia, Covidl9, Normal, and Viral Pneumonia). It can hardly distinguish between bacterial pneumonia and viral pneumonia. Based on CNN, transfer learning, and interpretable machine learning methods, this work precisely implements data processing and augmentation and adds a second binary classifier following a confidence level. In this way, the accuracy and recall rate of the quaternary classification are significantly improved, especially for bacterial pneumonia and viral pneumonia, and the model also becomes more interpretable.