Transfer and deep learning techniques for the automatic diagnosis of COVID-19 respiratory diseases

Abstract

The current coronavirus disease (COVID-19) outbreak has demonstrated the need to develop rapid and accurate diagnostic tools in order to assess the different types of respiratory diseases that affect COVID-19 patients that are significant and critical for the patient care and treatment. X-Ray imaging is one of the most important radiological examinations for screening and diagnosis of lung diseases. In this work, we describe a deep-learning based approach appropriate to potentially diagnose automatically patterns and characteristics from medical X-Ray images associated with known respiratory diseases as well as COVID-19 related ones. We propose a Convolutional Neural Network (CNN) framework for multi-label classification of the fourteen respiratory diseases and that of healthy patients. Here, we note the dataset shows disproportionate representations of the diseases. We have trained the CNN model on large multidisease, physician-diagnosed X-ray images available on an NIH open database source. We tested several loss functions commonly recommended for multi-classification training, and determined that Multi-Labeled Margin Soft loss function shows a rather smooth optimization with an apparent exponential behavior with training epoch. Following a transfer learning approach, we extend the parameters obtained from the training on the large data set to train and assess newly acquired X-ray images from COVID-19 infected patients but not labeled for any particular respiratory disease. Although additional work on validation and testing of the CNN model is needed, we have identified several parameters relevant for accurate classification