The 2019 novel coronavirus (COVID-19) is an infectious disease caused by coronavirus primarily affecting the respiratory tract. Coronaviruses (CoV) are a large family of viruses that can cause illnesses such as the common cold, severe acute respiratory syndrome (SARS), and Middle East respiratory syndrome (MERS). COVID-19 outbreak was first reported in Wuhan, China and has spread rapidly to other countries. As of 23rd June 2020, there are a total of 8,993,659 confirmed cases with 469,587 deaths in more than 227 countries across the globe [1]. The virus is now known as the severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2). In March 2020, the World Health Organization (WHO) declared the COVID-19 outbreak a pandemic.

Considering the fast spread of the disease and pandemic situation, fast testing, diagnosis and treatment of patients is highly desirable. The standard COVID-19 tests called PCR (Polymerase chain reaction) is available but has limitations. Pathogenic laboratory testing is the diagnostic gold standard for COVID, but it is time-consuming and has high false negatives [2]. Moreover, large scale implementation of the COVID-19 tests which are extremely expensive cannot be afforded by many of the developing and underdeveloped countries. Therefore, a need arises for an alternate automated system, facilitating diagnosis and testing procedures using artificial intelligence and machine learning.

Transfer Learning has become immensely useful in medical applications since it does not require as much training data, which can be hard to get in medical imaging use cases. As we have a relatively small COVID positive image dataset, transfer learning was an optimal choice for our experiment. Transfer learned models has been trained on an extremely large ImageNet dataset, and we can transfer weights which were learned through hundreds of hours of training on multiple high-powered graphics processing units (GPUs). The training process to update weights of the layers of deep convolutional neural network (DCNN) is otherwise very expensive to achieve from scratch [3]. Many such models are available as open-source and hence easy to access [4].

We have conducted a study leveraging transfer leaning networks of different DCNN models (ResNet50, InceptionV3 and VGG-16) to detect pneumonia infected patients using chest X-ray images. In this session we will present a brief overview of transfer leaning technique with comparative study of the model performance of these networks for COVID image classification using chest X-ray images. Also, we will showcase a “visual explanation” of COVID infection localization in infected lungs using gradient-based class activation maps (GRAD-CAM) [5]. Grad-CAM is a class-discriminative localization technique that generates visual explanations for any CNN-based network without requiring architectural changes or re-training.

References
1. World Health Organization; Coronavirus disease (COVID-2019) situation reports; https://www.who.int/emergencies/diseases/novel-coronavirus-2019/situation-reports.
2. Assay Techniques and Test Development for COVID-19 Diagnosis; Carter et al, ACS Cent Sci. 2020 May 27; 6(5): 591–605.
3. Convolutional Neural Networks for Medical Image Analysis: Full Training or Fine Tuning? Tajbakhsh et al, IEEE Trans Med Imaging. 2016 May;35(5):1299-1312.
4. https://keras.io/api/applications/
5. Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization; Selvaraju et al. arXiV; 2019 Dec; Computer Vision and Pattern Recognition.