Synthetic Lung Ultrasound Data Generation Using Autoencoder With Generative Adversarial Network.

Area of Science:

• Medical imaging
• Artificial intelligence
• Ultrasound technology

Background:

• Class imbalance poses a significant challenge in medical image analysis, especially in lung ultrasound (LUS) where severe patterns are underrepresented.
• Traditional oversampling methods often prove insufficient for effectively handling this data scarcity.

Purpose of the Study:

• To introduce a novel supervised autoencoder generative adversarial network (SA-GAN) for data augmentation in LUS analysis.
• To enhance the quality of synthetic samples for minority classes using generative AI.
• To compare the performance of SA-GAN augmentation against traditional techniques in multiclass classification tasks.

Main Methods:

• Development of a SA-GAN incorporating an autoencoder to create a conditional latent space, addressing weight clipping issues.
• Generation of high-quality synthetic LUS data for underrepresented classes.
• Evaluation of generated samples using similarity metrics and expert analysis.
• Comparison of state-of-the-art neural network performance trained with SA-GAN versus traditional augmentation.

Main Results:

• The SA-GAN effectively generates high-quality synthetic data, addressing limitations of traditional augmentation.
• Expert analysis and similarity metrics validate the utility of the generated samples.
• AI models trained with SA-GAN augmentation demonstrate improved robustness and reliability in LUS multiclass classification.

Conclusions:

• The proposed SA-GAN is a powerful tool for mitigating class imbalance in LUS analysis.
• This generative AI approach enhances the performance and reliability of diagnostic AI models.
• SA-GAN offers a promising solution for improving medical image analysis in scenarios with limited data.