Publication: COVID-19 identification based on CNN through transfer learning
Loading...
Date
2023-07
Authors
Ang, Zheng Xiong
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
The reported cases of COVID-19 disease are increasing drastically in the current days. Medical imaging techniques for lung disease diagnosis, such as CXR, are important in providing detailed information about the disease’s burden, especially for individuals experiencing moderate to severe COVID-19 infection. Apart from that, the workload of medical experts in examining the medical images also increased significantly during this COVID-19 pandemic. Due to several advantages of CXR as compared to MRI, CXR image dataset is used in this project to train and validate the CNN model via transfer learning in MATLAB. There are several factors which could affect the performance of the pre-trained networks: number of training dataset, and modification on training dataset such as masking and data augmentation. Therefore, ResNet-18 is trained with two different sets of datasets. The first set is the original CXR image without masking and
image augmentation, while another set is masked CXR image with image augmentation. Based on the result, the increase in training dataset has improved the accuracy of the CNN model from 86.69% to 96.17%. Besides, the CXR image dataset which has been processed with masking technique outperformed the original CXR image. Additionally, other pre-trained CNN models used in this project are GoogLeNet, ResNet-18, ResNet 50 and ResNet-101. By using confusion matrix, the performance of these different pre trained networks on CXR image classification is evaluated, ResNet-18 has achieved the highest accuracy of 96.17%. For future improvements, the pre-trained network which comes out with the highest accuracy can be further applied with more image datasets and
wider range of data augmentation to improve the performance of the model.