生成式文本到图像AI模型在整形美容外科中种族、性别和年龄多样性的探索
本研究评估了DALL-E3、Midjourney和Adobe Firefly三个生成式AI模型在抽脂、眼睑成形术和鼻成形术提示下生成图像的种族、性别和年龄多样性。结果显示模型间浅色皮肤与深色皮肤代表性无显著差异(P=0.26和P=0.31),但模型间及男女间在衰老表现上存在显著差异(P<0.0001和P=0.0009),在清透皮肤、大眼和/或浅色眼睛、窄鼻子的描绘上也有显著差异。结论认为浅色和深色皮肤代表性尚可,但性别偏见和西方化审美标准的表现有待改善。
要点
- 三个AI模型在浅色与深色皮肤代表性上无显著差异(P=0.26和P=0.31)。
- 模型间及男女间在衰老表现上存在显著差异(P<0.0001和P=0.0009)。
- 清透皮肤、大眼和/或浅色眼睛、窄鼻子的描绘有显著差异(P<0.0001、P=0.0010、P<0.0001)。
核心结论
- 尽管模型间浅色和深色皮肤代表性尚可,但性别偏见和西方化审美标准的表现有待改善。
研究设计与样本
- 研究使用DALL-E3、Midjourney和Adobe Firefly三个生成式文本到图像AI模型,针对抽脂、眼睑成形术和鼻成形术,按性别(男/女)、种族(白人、黑人、拉丁裔)和年龄组(20-30、31-45、46+岁)组合设计提示生成图像,并采用Fitzpatrick和Monk量表评估肤色、4项问卷评估性别对等性、以及西方化审美标准纳入情况,使用Kruskal-Wallis检验或Fisher精确检验进行模型间比较(P<0.05)。
主要发现
- 模型间浅色皮肤(Fitzpatrick I-III和Monk 1-5)与深色皮肤(Fitzpatrick IV-VI和Monk 6-10)代表性无显著差异(P=0.26和P=0.31)。
- 模型间及男女间在衰老表现上存在显著差异(P<0.0001和P=0.0009)。
- 清透皮肤(P<0.0001)、大和/或浅色眼睛(P=0.0010)、窄鼻子(P<0.0001)的描绘存在显著差异。
关键数值与推荐
- 肤色评估采用Fitzpatrick量表(I-III为浅色,IV-VI为深色)和Monk量表(1-5为浅色,6-10为深色)。
- 性别对等性使用4项问卷评估。
- 统计分析使用Kruskal-Wallis检验或Fisher精确检验,显著性水平P<0.05。
临床怎么用
- 临床使用生成式AI模型生成手术效果图像时,需注意其可能存在的性别偏见和西方化审美标准倾向。
局限性
- 原文未述局限性
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
BACKGROUND: It is unclear how representative and inclusive of various patient populations generative text-to-image artificial intelligence (AI) models are. OBJECTIVES: This project explores the diversity of race, gender, and age in the images generated by AI models: DALL-E3, Midjourney, and Adobe Firefly, in response to prompts focused on liposuction, blepharoplasty, and rhinoplasty. METHODS: Prompts were designed to prompt the AI model to generate images of surgical outcomes for liposuction, blepharoplasty, and rhinoplasty for each gender, race, and age combination: male vs female, Caucasian or White, Black or African American, Latino or Hispanic, and age groups: 20 to 30, 31 to 45, and 46+ years. Each generated image was evaluated for representation of skin color by Fitzpatrick and Monk scales, sex parity using a 4-item questionnaire, and the incorporation of westernized beauty standards. Analysis was then conducted, utilizing the Kruskal-Wallis test or the Fisher's exact test between the 3 models (P < 0.05). RESULTS: There was no significant difference between the representation of light skin color (Fitzpatrick I-III and Monk 1-5) vs dark skin color (Fitzpatrick IV-VI and Monk 6-10) between the models (P = 0.26 and P = 0.31). A significant difference was found between the models and between females vs males regarding aging (P < 0.0001 and P = 0.0009). There were also significant differences found for the depiction of clear skin (P < 0.0001), large and/or light-colored eyes (P = 0.0010), and narrow noses (P < 0.0001). CONCLUSIONS: Although there is fair representation of light skin colors and dark skin colors across the models, the depiction of gender bias and westernized beauty standards can be improved.
引用
- Trabilsy M, Genovese A, Prabha S, et al. Utilizing Generative Text-to-Image Artificial Intelligence Models to Explore Race, Gender, and Age in Plastic and Aesthetic Surgery. Aesthet Surg J. 2025. PMID: 40378272. doi:10.1093/asj/sjaf084
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。