Neural Style Transfer as Data Augmentation for Improving COVID-19 Diagnosis Classification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Coronavirus disease 2019 (COVID-19) has caused millions of deaths and strained healthcare systems.
• Accurate and accessible diagnostic tools are crucial for managing the pandemic, but current COVID-19 testing is insufficient and costly.
• Chest X-rays are vital for diagnosing respiratory illnesses, including COVID-19, but limited availability of positive COVID-19 X-ray datasets hinders AI model development.

Purpose of the Study:

• To address the scarcity of COVID-19 positive chest X-ray images for training diagnostic AI models.
• To evaluate the effectiveness of cycle-generative adversarial networks (CycleGAN) for augmenting limited COVID-19 positive X-ray datasets.
• To enhance the performance of convolutional neural networks (CNNs) in classifying COVID-19 from chest X-rays through data augmentation.

Main Methods:

• Utilized cycle-generative adversarial networks (CycleGAN), a technique from neural style transfer, to synthesize realistic COVID-19 positive X-ray images from COVID-19 negative images.
• Augmented a dataset of chest X-rays by generating synthetic COVID-19 positive samples.
• Integrated the augmented dataset with standard transfer learning techniques to train various convolutional neural networks (CNNs).

Main Results:

• Demonstrated a significant increase in the mean macro F1-score by over 21% for COVID-19 classification.
• Achieved statistical significance with a one-tailed t-score of 2.68 and a p-value of 0.01, supporting the effectiveness of the augmentation method.
• Showcased improved performance of common CNN architectures when trained on the augmented dataset.

Conclusions:

• Cycle-generative adversarial networks are effective for augmenting limited COVID-19 positive chest X-ray datasets.
• This data augmentation approach, combined with transfer learning, substantially improves the diagnostic performance of AI classifiers for COVID-19 detection.
• The method offers a viable solution to overcome data scarcity challenges in developing robust AI tools for infectious disease diagnosis.