Deep Convolutional Neural Network (CNN) Design for Pathology Detection of COVID-19 in Chest X-Ray Images

• Published in: Machine Learning and Knowledge Extraction Vol. 12844; pp. 211 - 223
• Main Authors: Darapaneni, Narayana, Sil, Anindya, Kagiti, Balaji, Krishna Kumar, S., Ramanathan, N. B., VasanthaKumara, S. B., Paduri, Anwesh Reddy, Manuf, Abdul
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer International Publishing AG 2021; Springer International Publishing
• Series: Lecture Notes in Computer Science
• Subjects: CNN; COVID-19; DarkNet-53; InceptionNet-V3; McNemar-Bowker test; MobileNet-V2; ResNet-50; Transfer learning; VGG-16
• ISBN: 303084059X; 9783030840594
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-030-84060-0_14

The coronavirus disease 2019 (COVID-19) caused by a novel coronavirus, turned into a pandemic and raised a serious concern to the global healthcare system. The reverse transcription polymerase chain reaction (RT-PCR) is the most widely used diagnostic tool to detect COVID-19. However, this test is time consuming and subject to availability of the test kits during a crisis. An automated method of screening chest x-ray images using convolutional neural network (CNN) Transfer Learning approach has been proposed as a relatively fast and cost-effective, decision support tool to detect pulmonary pathology due to COVID-19. In this study we have used Kaggle dataset with chest x-ray images of normal and pneumonia cases. We have added COVID-19 x-ray images from 5 different open-source datasets. The images were pre-processed based on the position of radiography images and greyscale was applied and subsequently the images were used for training. After consolidation, COVID-19 images comprised only 5% of the dataset. To address the class imbalance, we have used dynamic image augmentation technique to reduce the bias. We have then explored custom CNN and VGG-16, InceptionNet-V3, MobileNet-V2, ResNet-50, and DarkNet-53 transfer learning approaches to classify COVID-19, other pneumonia and normal x-ray images and compared their performances. So far, we have achieved the best score of F1 score 0.95, sensitivity 95% and specificity 95% for COVID-19 class with Darknet-53 feature extractor. Darknet-53 classifier is part of the state-of-the-art object detection algorithm named Yolo-v3. We have also done a McNemar-Bowker post-hoc test to compare Darknet-53 performance with the next best ResNet-50. This test suggests that Darknet-53 is significantly better skilled than ResNet-50 in differentiating COVID-19 from other pneumonia in chest x-ray images.