Identifying Obviously Artificial Medical Images Produced by a Generative Adversarial Network

Area of Science:

• Medical image processing
• Artificial intelligence
• Generative models

Background:

• Synthetic medical images are crucial for developing AI systems, especially in niche domains with limited data.
• Current evaluation metrics for synthetic images (e.g., Visual Turing Test, Fréchet Inception Distance) assess group-level quality, not individual image artificiality.
• Poorly generated synthetic images can negatively impact the performance of AI systems that assimilate them.

Purpose of the Study:

• To develop an automated method for identifying obviously artificial synthetic medical images.
• To filter out low-quality synthetic images that could compromise AI system performance.
• To establish a quality control measure for synthetic medical image datasets.

Main Methods:

• Synthetic computed tomography (CT) images generated by a progressively-grown generative adversarial network (PGGAN) were used.
• Images were evaluated using the Visual Turing Test (VTT), and image embeddings were analyzed for correlation with artificiality.
• A support vector machine classifier was trained to identify images with a high probability (≥0.7) of being rated as fake.

Main Results:

• The support vector machine classifier achieved 75.5% accuracy in identifying obviously artificial synthetic CT images.
• The classifier demonstrated a sensitivity of 0.743 and a specificity of 0.769.
• These results indicate a promising approach for validating synthetic medical image datasets.

Conclusions:

• The developed method offers a potential solution for quality control of synthetic medical images.
• Automated detection of artificial images is essential before their deployment in next-generation medical AI systems.
• This approach contributes to ensuring the reliability and performance of AI-driven medical image processing.