Some radiologists may not be able to reliably determine if images were generated using AI, according to research published July 16 in Clinical Radiology.
In a study, radiologists correctly identified AI-generated images about 75% of the time, with CT and MR images being harder to distinguish, wrote Robert Cronshaw, PhD, and Michelle Williams, PhD, from the University of Edinburgh in Scotland.
“These results suggest that AI generated radiological images could be successfully used for tasks such as model training, but it also raises concerns about the potential for misuse of these tools for malicious purposes,” Cronshaw and Williams wrote.
Synthetic images can be created with a text prompt and iterative denoising. Diffusion models can produce a diverse set of output images based on input text prompts, with domain-specific fine tuning improving image quality.
The researchers explored whether radiologists could distinguish between real radiological images and AI-generated images. They assessed synthetic radiological images created by using a fine-tuned stable diffusion model.
Top eight incorrectly classified synthetic images (Most convincing fake images). Images A, B, C, D, E, F, G and H were the most infrequently correctly classified images at 24%, 43%, 59%, 64%, 69%, 71%, 71% and 73% respectively. All images are AI-generated. Images A, B and C are head CTs. Images D and E are chest CTs. Image F is a cardiac CT. Image G is a musculoskeletal MRI. Image H is a head MRI. Images are republished under a Creative Commons license (CC BY 4.0).
The study included 10 real images and 20 AI-generated images. The researchers asked 182 radiologists about their confidence in their decisions on a one-to-five scale (five being the highest).
They reported the following findings:
-
Radiologists correctly identified AI-generated images 75.0% of the time compared to 83.4% of real images (p = 0.19).
-
Radiologists were more likely to correctly identify ultrasound (88%) and x-ray (91%) than cross sectional images like CT (70%) or MRI (77%). However, these differences did not achieve statistical significance (p = 0.015).
-
Radiologists reported having similar confidence for AI-generated and real images (3.50 vs. 3.56, p = 0.49).
Finally, the researchers observed no difference in classification based on number of years of experience (p = 0.57) or familiarity with AI (p = 0.37). However, radiologists with relevant specialist interests were more likely to correctly classify images (80.7% vs. 76.9%, p = 0.012).
Cronshaw and Williams highlighted that synthetic images could help improve AI model training, address privacy concerns, reduce biases in datasets, and be used in education. However, they acknowledged that their results “also raise concerns about the potential for misuse of these tools for malicious purposes.”
They called for reviewers and editors of medical journals to be aware of the risk of fake images in medical literature and to question any discrepancies.
“Fake medical images could also appear in patient records, which could lead to misdiagnosis and inappropriate treatments,” they wrote.
Read the full study here.
Leave a Reply