Overcoming barriers to data sharing with medical image generation: a comprehensive evaluation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Data Privacy

Background:

• Sharing patient data for medical research is hindered by privacy concerns.
• Researchers often need cohort-level insights, not individual patient information.

Purpose of the Study:

• To utilize generative adversarial networks (GANs) for creating synthetic medical imaging datasets.
• To assess the quality and utility of GAN-generated synthetic data for medical research.

Main Methods:

• Generated synthetic chest radiographs and brain CT scans using two GAN models.
• Evaluated synthetic data quality by training predictive models and comparing their performance on real vs. synthetic data.
• Conducted a reader study with radiologists to differentiate between real and synthetic images.

Main Results:

• Synthetic data performance improved with fewer classes and higher resolution.
• Label overfitting was observed in GAN training with low samples per class.
• Radiologist classification accuracy increased with higher image resolution.

Conclusions:

• Synthetic medical data generated by GANs can be a viable alternative to real patient data in specific research contexts.
• Guidelines for effective synthetic data generation and application in medical research were provided.
• GAN-generated data offers a promising solution for privacy-preserving medical data sharing.