医学成像中图像生成式人工智能的性别和种族偏见，第2部分：DALL-E 3分析

Currie Geoffrey, Hewis Johnathan, Hawk Elizabeth, Rohren Eric

Charles Sturt University, Wagga Wagga, New South Wales, Australia;

Baylor College of Medicine, Houston, Texas.

J Nucl Med Technol. 2024 Oct 22. doi: 10.2967/jnmt.124.268359.

Disparity among gender and ethnicity remains an issue across medicine and health science. Only 26%-35% of trainee radiologists are female, despite more than 50% of medical students' being female. Similar gender disparities are evident across the medical imaging professions. Generative artificial intelligence text-to-image production could reinforce or amplify gender biases. In March 2024, DALL-E 3 was utilized via GPT-4 to generate a series of individual and group images of medical imaging professionals: radiologist, nuclear medicine physician, radiographer, nuclear medicine technologist, medical physicist, radiopharmacist, and medical imaging nurse. Multiple iterations of images were generated using a variety of prompts. Collectively, 120 images were produced for evaluation of 524 characters. All images were independently analyzed by 3 expert reviewers from medical imaging professions for apparent gender and skin tone. Collectively (individual and group images), 57.4% ( = 301) of medical imaging professionals were depicted as male, 42.4% ( = 222) as female, and 91.2% ( = 478) as having a light skin tone. The male gender representation was 65% for radiologists, 62% for nuclear medicine physicians, 52% for radiographers, 56% for nuclear medicine technologists, 62% for medical physicists, 53% for radiopharmacists, and 26% for medical imaging nurses. For all professions, this overrepresents men compared with women. There was no representation of persons with a disability. This evaluation reveals a significant overrepresentation of the male gender associated with generative artificial intelligence text-to-image production using DALL-E 3 across the medical imaging professions. Generated images have a disproportionately high representation of white men, which is not representative of the diversity of the medical imaging professions.

性别和种族差异在医学和健康科学领域仍然是一个问题。尽管超过50%的医学生是女性，但实习放射科医生中只有26%-35%是女性。类似的性别差异在医学影像专业中也很明显。生成式人工智能文本到图像的生成可能会强化或放大性别偏见。2024年3月，通过GPT-4使用DALL-E 3生成了一系列医学影像专业人员的个人和群体图像：放射科医生、核医学医生、放射技师、核医学技术人员、医学物理学家、放射药剂师和医学影像护士。使用各种提示生成了多轮图像。总共生成了120张图像，用于评估524个角色。所有图像均由3名医学影像专业的专家评审员独立分析，以确定明显的性别和肤色。总体而言（个人和群体图像），57.4%（=301）的医学影像专业人员被描绘为男性，42.4%（=222）为女性，91.2%（=478）为浅肤色。放射科医生的男性比例为65%，核医学医生为62%，放射技师为52%，核医学技术人员为56%，医学物理学家为62%，放射药剂师为53%，医学影像护士为26%。与女性相比，所有职业中男性的比例都过高。没有残疾人士的形象。这项评估揭示了在医学影像专业中，使用DALL-E 3进行生成式人工智能文本到图像生成时，男性的比例明显过高。生成的图像中白人男性的比例过高，这并不代表医学影像专业的多样性。