Application of Generative Adversarial Networks to Improve COVID-19 Classification on Ultrasound Images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Lung ultrasound is a cost-effective tool for COVID-19 screening, but its interpretation requires expertise.
• Deep learning models can automate COVID-19 classification from ultrasound, yet face challenges due to limited public data.
• Data scarcity hinders the development of robust deep learning models for medical image analysis.

Purpose of the Study:

• To address data scarcity in COVID-19 lung ultrasound screening by generating synthetic images using GANs.
• To evaluate the effectiveness of synthetic data generated by Wasserstein GANs (WGAN) and Pix2Pix for training classification models.
• To improve the accuracy and generalizability of deep learning models for COVID-19 diagnosis using augmented datasets.

Main Methods:

• Utilized Wasserstein GANs (WGAN) and Pix2Pix to generate synthetic lung ultrasound images.
• Trained frame-based classification models using both real and GAN-generated synthetic data.
• Validated the quality of synthetic data by comparing its distribution to the original dataset using specific analytical tools.

Main Results:

• Synthetic data generated by GANs exhibited a distribution closely resembling the original dataset.
• Classification models trained with synthetic data achieved a peak accuracy of 96.32% ± 4.17%.
• Models trained exclusively on real data reached a maximum accuracy of 82.69% ± 10.42%, indicating a significant performance improvement with synthetic data.

Conclusions:

• GAN-generated synthetic lung ultrasound data effectively mitigates data scarcity issues in medical AI.
• The proposed method significantly enhances the performance of deep learning models for COVID-19 screening.
• This approach offers a promising avenue for developing more accurate and accessible AI-driven diagnostic tools in medical imaging.