Real or not real? Can radiologists distinguish artificial intelligence generated radiological images from real ones?
Authors
Affiliations (2)
Affiliations (2)
- British Heart Foundation Centre for Research Excellence, Institute for Neuroscience and Cardiovascular Research, University of Edinburgh, Chancellor's Building, 49 Little France Crescent, Edinburgh EH16 4TJ, United Kingdom.
- British Heart Foundation Centre for Research Excellence, Institute for Neuroscience and Cardiovascular Research, University of Edinburgh, Chancellor's Building, 49 Little France Crescent, Edinburgh EH16 4TJ, United Kingdom. Electronic address: [email protected].
Abstract
Artificial intelligence (AI) models can create radiological images. We aimed to determine whether radiologists could distinguish AI-generated from real images, and factors associated with correct classification. AI-generated images were made using an implementation of the Dreambooth fine-tuning approach applied to Stable Diffusion v2.1. Radiologists were asked to classify images as real (n = 10) or AI-generated (n = 20) and their confidence in this decision (1 least, 5 most) in an online form. 182 radiologists completed the survey. The median proportion of correctly identified images per respondent was 77.8% (interquartile range, IQR 70.0, 86.7%), with no difference between AI-generated (75.0%, IQR 70.5, 87.1%) and real images (83.4%, IQR 74.2, 92.6%, p = 0.19). Ultrasound and X-ray were more likely to be correctly identified than cross-sectional images like CT or MRI (88%, 91%, 70% and 77% respectively, p = 0.015). Mean confidence was similar for AI-generated and real images (3.50 ± 0.23 versus 3.56 ± 0.23, p = 0.49). There was 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 ± 1.3% versus 76.9 ± 0.8%, p = 0.012). Radiologists were only able to correctly identify three-quarters of AI-generated images. This was impacted by sub-specialist expertise but not the number of years of experience or familiarity with AI.