Deepfake X-Rays Deceive Radiologists and AI Models

A study published in Radiology reveals that both radiologists and multimodal large language models (LLMs) struggle to distinguish AI-generated deepfake X-ray images from authentic ones. The study involved 17 radiologists from 12 centers across six countries, who evaluated 264 X-ray images, half of which were real and half AI-generated. When unaware of the study's purpose, only 41% of radiologists identified the AI-generated images, and their mean accuracy improved to 75% after being informed. The research highlights significant cybersecurity risks, as deepfake images could lead to fraudulent litigation and compromised patient diagnoses if hackers manipulate medical records. The findings emphasize the urgent need for tools and training to help healthcare professionals detect deepfakes. The study also assessed the performance of four LLMs, which had accuracy rates ranging from 57% to 85% in identifying the synthetic images.

Key Points: • Radiologists and AI models struggle to identify deepfake X-rays, with only 41% accuracy initially. • The study involved 17 radiologists analyzing 264 X-ray images, half of which were AI-generated. • There are significant cybersecurity risks associated with deepfake medical images affecting patient diagnoses.