A Critical Assessment of Generative Models for Synthetic Data Augmentation on Limited Pneumonia X-ray Data.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Pathology

Background:

• Deep learning models require extensive data for medical image analysis.
• Limited data in medical imaging hinders the training of effective diagnostic models.
• Artificial image generation is a potential solution to data scarcity.

Purpose of the Study:

• To compare five generative models for synthetic medical image creation.
• To evaluate the impact of synthetic data on a pneumonia classification task.
• To assess the relationship between synthetic image quality and classification performance.

Main Methods:

• Trained five generative models, including Generative Adversarial Networks (GANs) and diffusion models.
• Evaluated synthetic data on a downstream classification task for pneumonia detection using 1082 chest X-ray images.
• Assessed image quality, pathological plausibility, and classification performance using various metrics.

Main Results:

• GAN-based approaches produced higher quality and more plausible synthetic images than diffusion models.
• Improved image quality did not consistently translate to better classification performance.
• Synthetic data improved class-specific metrics, particularly for underrepresented classes, rebalancing the dataset.
• One model, DreamBooth, showed a slight improvement in overall accuracy (+0.52).

Conclusions:

• Generative models can enhance data availability in limited medical imaging scenarios.
• The choice of generative model impacts synthetic data quality and downstream task performance.
• Careful consideration is needed when applying generative models, as image quality doesn't always correlate with improved classification outcomes.