AUTHOR=Pannipulath Venugopal Vishnu , Babu Saheer Lakshmi , Maktabdar Oghaz Mahdi 

TITLE=COVID-19 lateral flow test image classification using deep CNN and StyleGAN2

JOURNAL=Frontiers in Artificial Intelligence

VOLUME=Volume 6 - 2023

YEAR=2024

URL=https://www.frontiersin.org/journals/artificial-intelligence/articles/10.3389/frai.2023.1235204

DOI=10.3389/frai.2023.1235204

ISSN=2624-8212

ABSTRACT=The adoption of artificial intelligence in the healthcare industry has the potential to significantly improve clinical workflows and diagnoses. This can be particularly beneficial for time-critical large-scale operations, as exemplified by recent mass testing efforts for COVID-19 and the accurate identification and tracking of positive patients. This study presents an automated COVID-19 Rapid Antigen Test Device (RATD) image classification model that relies on deep Convolutional Neural Network (CNN) to differentiate between positive and negative real-world RATD test images. Such a system can greatly expedite the testing process and facilitate a large-scale testing and tracking system. One major challenge in the design and development of such a model is the absence of RATD image dataset for training deep CNN classification models. To address this issue, this study compiled a dataset of RATD images through crowdsourcing using mobile devices. A total of 900 real-world images were captured, with a focus on two specific classes of images: Positive (COVID-19 detected) and Negative (COVID-19 not detected). These images were used to develop a binary CNN classifier. Rigorous data augmentation was performed to further increase the dataset's size. Moreover, StyleGAN2-ADA was used to generate simulated training sample images, primarily to address issues related to dataset class imbalance. The performance of several classification CNN models was investigated, with 1 Venugopal et al.the best model achieving a validation accuracy of 93% and test accuracies of 88% and 82% for simulated and real datasets, respectively. It was observed that further augmenting simulated images during training did not significantly improve the performance of realworld test images; instead, it only improved the performance of simulated test images.The findings of this study can be applied in the deployment of large-scale test and tracking systems to mitigate outbreaks similar to COVID-19.