VGG-FusionNet: A Feature Fusion Framework from CT scan and Chest X-ray Images based Deep Learning for COVID-19 Detection

ABSTRACT

COVID-19 has been rampant across the globe since it was discovered in 2020, but the method of virus detection still lacks efficiency and requires human resources. Given the slow delivery of the PCR test and the many possible false negatives of the rapid tests, medical imaging such as a chest computed tomography (CT) scan or chest X-ray (CXR) is an alternative and efficient way to detect the coronavirus accurately. For the past two years, many researchers have proposed different deep learning methods for COVID-19 detection using CT scans or CXR images. Due to the lack of available data, our study aims to propose a new deep learning framework VGG-FusionNet that takes advantage of integrating features from both CT scan and CXR images while avoiding some pitfalls from previous studies, including a high risk of bias due to lack of demographic information for the dataset, poor reproducibility, and no evaluation on different data sources to study the generalizability. Specifically, we use the convolutional layers of GoogLeNet, ResNet, and VGG to extract features from CT scan and CXR images and fuse them before training through fully connected layers. The result shows that using VGG's convolutional layers achieves the best overall performance with an accuracy of 0.93. Our proposed framework outperforms the deep learning models, using features from CT scans or CXR. © 2022 IEEE.