Publication: Deep Learning Approaches for Automated Diagnosis of COVID-19 Using Imbalanced Training CXR Data
Loading...
Date
2022
Authors
Journal Title
Communications in Computer and Information Science
Journal ISSN
Volume Title
Publisher
Springer Science and Business Media Deutschland GmbH
Abstract
Due to the exponential rise of COVID-19 worldwide, it is important that the artificial intelligence community address to analyze CXR images for early classification of COVID-19 patients. Unfortunately, it is very difficult to collect data in such epidemic situations, which is essential for better training of deep convolutional neural networks. To address the limited dataset challenge, the author makes use of a deep transfer learning approach. The presence of limited number of COVID-19 samples may lead to biased learning due to class imbalance. To resolve class imbalance, we propose a new class weighted loss function that reduces biasness and improves COVID-19 sensitivity. Classification and preprocessing are two concrete components of this study. For classification, we compare five pre-trained deep neural networks architectures i.e. DenseNet169,�InceptionResNetV2, MobileNet, Vgg19 and�NASNetMobile�as a baseline to achieve transfer learning. This study is conducted using two fused datasets where samples are collected from four heterogeneous data resources. Based on number of classes we make four different classification scenarios to compare five baseline architectures in two stages. These scenarios are COVID-19 vs non- COVID-19, COVID-19 vs Pneumonia vs Normal, COVID-19 pneumonia vs Viral vs Bacterial pneumonia vs Normal and COVID19 vs Normal vs Virus + Bacterial pneumonia. The primary goal of this study is to improve COVID-19 sensitivity. Experimental outcomes show that DenseNet169 achieves the highest accuracy and sensitivity for COVID-19 detection with score of 95.04% and 100% for 4-class classification and 99.17% and 100% for 3 class-classification. � 2022, Springer Nature Switzerland AG.
Description
Keywords
Chest X-ray, COVID-19 classification, Deep learning, Image processing, Imbalanced data, Pneumonia classification