To evaluate the detectability and reader-perceived visual authenticity of synthetic chest radiographs generated by the GPT-image and Gemini-image models.
Approach:
Study Design: A blinded single-image reader study was conducted with 320 real disease-positive frontal chest radiographs and 320 matched normal conditioning radiographs.
Generation Strategies: Two generation strategies were compared: text-only generation and image-conditioned generation based on normal conditioning radiographs.
Assessment Method: Four radiologists independently assessed the synthetic and real disease-positive radiographs, followed by a paired evaluation of previously undetected image-conditioned synthetic radiographs.
Image Similarity Metrics: Image-level similarity was assessed using SSIM, LPIPS, and FID.
Key Findings:
Image-conditioned generation had significantly lower AI detection rates than text-only generation (34.0% vs. 56.1%).
Synthetic radiographs generated by the GPT-image model were less frequently detected than those generated by the Gemini-image model.
Synthetic radiographs showed higher image-level structural and perceptual similarity to the conditioning radiographs.
Paired comparison improved detection of previously undetected image-conditioned synthetic radiographs.
Interpretation:
Image-conditioned GenAI can produce synthetic chest radiographs that may appear visually authentic to radiologists.
Limitations:
The study was retrospective and may not reflect real-world clinical settings.
Only two generative models were evaluated, limiting generalizability.
Conclusion:
The findings highlight the importance of evaluating the visual realism and detectability of synthetic medical images before their integration into healthcare and education.
