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

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Infectious Disease Diagnostics

Background:

• Artificial intelligence (AI) offers potential to improve healthcare workflows and diagnostic accuracy, especially for large-scale public health initiatives like COVID-19 testing.
• Automated image classification is crucial for efficient and accurate interpretation of diagnostic tests.

Purpose of the Study:

• To develop and evaluate a deep Convolutional Neural Network (CNN) model for automated classification of COVID-19 RATD images.
• To address limitations in available RATD image datasets through data augmentation and synthetic data generation.

Main Methods:

• A dataset of 900 real-world COVID-19 RATD images was crowdsourced.
• Data augmentation techniques and StyleGAN2-ADA were employed to generate synthetic images, mitigating dataset limitations and class imbalance.
• A deep CNN model was trained and validated on both real and synthetic datasets.

Main Results:

• The best performing CNN model achieved 93% validation accuracy.
• On test datasets, the model attained 88% accuracy with simulated images and 82% accuracy with real-world images.
• While data augmentation improved performance on simulated images, it did not significantly enhance accuracy on real-world test data.

Conclusions:

• The developed AI model demonstrates significant potential for expediting COVID-19 testing and supporting large-scale testing and tracking systems.
• Addressing dataset limitations and class imbalances is critical for the successful development of AI diagnostic tools.
• This research provides valuable insights for applying AI in mitigating future infectious disease outbreaks.