To assess the ability of radiologists to correctly identify artificial intelligence-generated radiological images compared to real images.
Approach:
Survey Methodology: 182 radiologists reviewed 30 images (20 AI-generated and 10 real) across various imaging modalities and subspecialties.
Image Generation: AI images were produced using Stable Diffusion version 2.1 and the DreamBooth approach, trained on 20 to 30 example images per modality.
Assessment Criteria: Correct classification of images as real or AI-generated was the primary assessment, with additional evaluations based on modality, experience, and confidence.
Key Findings:
Median correct classification rate was 78% across respondents.
75% of AI-generated images (2,640 assessments) and 83% of real images (1,820 assessments) were correctly identified.
Classification accuracy varied by imaging modality: CT (70%), MRI (77%), ultrasound (88%), and radiographs (91%).
Radiologists with relevant subspecialty expertise had higher accuracy (81%) compared to those without (77%).
Confidence in classification was correlated with correct identification, varying by imaging modality.
Interpretation:
Radiologists demonstrated variable accuracy in identifying AI-generated images, influenced by imaging modality and subspecialty expertise.
Limitations:
Images were presented online, limiting interaction compared to clinical practice.
Unequal numbers of AI-generated and real images across modalities may affect results.
Participants were aware that some images were AI-generated, potentially biasing their assessments.
AI models were trained primarily on normal anatomy, limiting generalizability to pathological images.
Potential selection bias due to participant experience and subspecialty.
Conclusion:
Radiologists may struggle to distinguish AI-generated images from real ones, with performance varying by modality and expertise.
