A recent study published in Radiology indicates that both radiologists and multimodal large language models (LLMs) struggle to differentiate between authentic X-ray images and those generated by artificial intelligence (AI). This research underscores the potential risks associated with AI-generated medical images and emphasizes the necessity for enhanced tools and training to maintain the integrity of medical imaging.
Key Findings:
- Seventeen radiologists from twelve institutions across six countries participated in the study.
- The study analyzed a total of 264 X-ray images, split evenly between real and AI-generated images.
- When unaware of the presence of synthetic images, only 41% of radiologists identified the AI-generated X-rays.
- After being informed about the synthetic images, the average accuracy improved to 75%.
- Individual performance varied, with accuracy ranging from 58% to 92%.
- Four LLMs, including GPT-4o and GPT-5, showed accuracy rates between 57% and 85% in detecting deepfake images.
Risks and Implications:
Lead author Mickael Tordjman, MD, highlighted the significant vulnerabilities posed by deepfake X-rays, particularly in legal contexts where fabricated images could lead to fraudulent claims. Additionally, there are cybersecurity concerns if malicious actors gain access to hospital networks and manipulate patient diagnoses.
Visual Clues in Deepfake X-Rays:
Researchers noted that deepfake medical images often exhibit unrealistic characteristics, such as:
- Overly smooth bones
- Unnaturally straight spines
- Excessively symmetrical lungs
- Uniform blood vessel patterns
- Clean and consistent fractures
Recommendations for Improvement:
To mitigate the risks associated with deepfake images, experts recommend:
- Implementing advanced digital safeguards, such as invisible watermarks.
- Embedding cryptographic signatures linked to technologists at the time of image capture.
Dr. Tordjman emphasized the importance of developing educational datasets and detection tools as AI technology continues to evolve, particularly with the potential for synthetic 3D imaging in the future.