Artificial intelligence (AI) has revolutionized art and design industries due to its ability to create images from natural language text. Such models also contain latent medical information. Text-to-image AI thus has the potential to create synthetic data sets for research, education, and communication. However, these may be indistinguishable from real images, causing issues with trust and potential misrepresentation. Radiologists and clinicians must be aware of the feasibility of creating “deep fake” medical images (Figure). Inconsistencies in anatomy or image texture could identify AI images, but there are currently no technical solutions for identification.
Cardiac CT images created using a subscription-based online text-to-image artificial intelligence (AI) system (DALL.E2, OpenAI; https://labs.openai.com). This deep learning AI was trained with text-image pairs from the internet. It generates images using an iterative process that starts with random dots. The text inputs were (A) “computed tomography of the heart,” (B) “computed tomography of the heart, greyscale,” and (C) “computed tomography of the heart with lots of colors.”
Footnotes
M.C.W. (FS/ICRF/20/26002), S.E.W. (FS/20/26/34952), and D.E.N. (CH/09/002, RG/16/10/32375, RE/18/5/34216) are supported by the British Heart Foundation. D.E.N. is supported by the Wellcome Trust (WT103782AIA).
Disclosures of conflicts of interest: M.C.W. Grant support from the British Heart Foundation (FS/ICRF/20/26002); speaker at meetings sponsored by Canon Medical Systems, Siemens Healthineers, and Novartis; Society of Cardiovascular Computed Tomography board of directors; member of the European Society of Cardiovascular Radiology Executive committee; British Society of Cardiovascular Imaging President-elect; Radiology: Cardiothoracic Imaging editorial board member. S.E.W. Grant support from the British Heart Foundation (FS/20/26/34952). D.E.N. Grant support from the Wellcome Trust (WT103782AIA).
Keywords: Cardiac, CT Angiography, AI-created Cardiac CT