Wasserstein GAN based Chest X-Ray Dataset Augmentation for Deep Learning Models: COVID-19 Detection Use-Case

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Infectious Diseases

Background:

• COVID-19 remains a global health concern, necessitating rapid and accurate diagnostic tools.
• Current RT-PCR tests have limitations, driving research into alternative methods like AI-based analysis of Chest X-rays (CXRs).
• Deep learning models show promise for COVID-19 detection from CXRs, but suffer from limited training data.

Purpose of the Study:

• To address the data deficiency challenge for deep learning models in COVID-19 detection using Chest X-ray images.
• To evaluate the effectiveness of Wasserstein Generative Adversarial Networks (WGANs) in generating synthetic COVID-19 CXR data.
• To demonstrate that WGAN-generated images are suitable for training and validating diagnostic AI models.

Main Methods:

• Utilized a Wasserstein Generative Adversarial Network (WGAN) to generate synthetic Chest X-ray images.
• Compared the quality and utility of WGAN-generated COVID-19 CXR images against original datasets.
• Performed inference tests using a pre-existing COVID-19 detection model on both real and generated CXR images.

Main Results:

• The WGAN successfully generated a dataset of COVID-19 Chest X-ray images.
• Generated images were found to be comparable in quality to original CXR images.
• Inference tests confirmed that WGAN-generated images maintain diagnostic performance when used with existing deep learning models.

Conclusions:

• Wasserstein Generative Adversarial Networks (WGANs) offer an effective and lightweight solution for augmenting limited COVID-19 CXR datasets.
• Synthetic data generated by WGANs can be reliably used to train and test deep learning models for COVID-19 detection.
• This approach helps overcome data scarcity, potentially improving the accessibility and performance of AI-driven COVID-19 diagnostics.