Diversity bias of text-to-image generative artificial intelligence in radiology and radiography

Introduction: Gender and ethnicity disparity remain an issue across medicine and healthcare 1 2. More than 50%3 of medical students are female, yet only 264 and 35%3 of trainee radiologists are women in Australia &amp; the USA respectively. In the USA, 51.4%5 of radiographers are female which is substantially lower than the 68.4% female workforce reported by the MRPBA in Australia6. Generative artificial intelligence (AI) text-to-image has a variety of applications and can create images representing radiography and radiology for education, recruitment, patients and the public7; however generative AI can reinforce or amplify gender and diversity biases8.<br/><br/>Method: DALL-E 3 was utilised via GPT-4 to generate a series of individual and group images of radiologists and radiographers. Multiple iterations of images were generated using a variety of prompts. All images were independently analysed by three expert reviewers for apparent gender and skin tone9. Ethics was not required. <br/><br/>Results: The male gender representation was 67.9% (N=57) for radiologists and 54.7% (N=41) for radiographers (Table 1). Light skin tone was represented for 92.9% (N=78) of radiologists and 92.0% (N=69) of radiographers. This represents over-representation of males compared to females for both professions. There was no representation of persons with a disability. <br/><br/>Conclusion: This evaluation reveals a significant male gender bias associated with generative AI text-to-image generation using DALL-E 3 for radiologists and radiographers. Generated images have a disproportionately high representation of white males which is not representative of the diversity of the professions with potential risk to amplify historical and institutionalised biases. <br/><br/>References:<br/><br/>1. Crews DC, Collins CA, Cooper LA. Distinguishing workforce diversity from health equity efforts in medicine. JAMA Health Forum. 2021;2(12):e214820. <br/>2. Lee R. How should medical school admissions drive health care workforce diversity? AMA J Ethics. 2021;23(12):E912-918.<br/>3. Lim W, Gupta N, Mandel C. Analysis of gender in radiology in Australia and its importance to the profession and workforce planning. J Med Imaging Radiat Oncol. 2023;67(2):170-178. <br/>4. Klontzas ME, Jean J, Turner VL, Balthazar P. Why and how to increase diversity in the radiology trainee workforce. Radiographics. 2022;42(3):E82-E85.<br/>5. Kubik-Huch RA, Vilgrain V, Krestin GP, Reiser MF, Attenberger UI, Muellner AU, Hess CP, Hricak H. Women in radiology: gender diversity is not a metric-it is a tool for excellence. Eur Radiol. 2020;30(3):1644-1652. <br/>6. MRPBA 2022. Workforce snapshot. https://www.ahpra.gov.au/documents/default.aspx?record=WD23%2f32488&amp;dbid=AP&amp;chksum=xUFAqrYOj6oTWjw6m6rk4A%3d%3d&amp;_gl=1*l4akmd*_ga*MTk2NzYzNzA3Mi4xNzA4NTEwNDYx*_ga_F1G6LRCHZB*MTcxMDgxMTUyMi40LjEuMTcxMDgxMTU3Ni4wLjAuMA.. Accessed 17 March 2024.<br/>7. Currie G, Hawk KE and Rohren E (2020). Ethical principles for the application of artificial intelligence (AI) in nuclear medicine and molecular imaging. European Journal of Nuclear Medicine and Molecular Imaging 47: 748–752.<br/>8. Currie G, Currie J, Anderson S, &amp; Hewis, J (2024). Gender bias in generative artificial intelligence text-to-image depiction of medical students. Health Education Journal 00:0;1-15.<br/>9. Massey DS, Martin JA. The NIS Skin Color Scale. Princeton, NJ: Princeton University Press. 2003.<br/>