Shiraishi Makoto, Banda Chihena Hansini, Nakajima Mayuri, Nakazwe Mildred, Wong Zi Yi, Tomioka Yoko, Moriwaki Yuta, Takeishi Hakuba, Lee Haesu, Kurita Daichi, Furuse Kiichi, Ohba Jun, Fujisawa Kou, Miyamoto Shimpei, Okazaki Mutsumi
Department of Plastic and Reconstructive Surgery, The University of Tokyo Hospital, Tokyo, Japan.
Plastic and Reconstructive Surgery Unit, Department of Surgery, The University Teaching Hospital, Lusaka, Zambia.
J Hand Microsurg. 2024 Nov 30;17(1):100196. doi: 10.1016/j.jham.2024.100196. eCollection 2025 Jan.
Since the release of ChatGPT by OpenAI in November 2022, generative artificial intelligence (AI) models have attracted significant attention in various fields, including surgery. These advancements have been particularly notable for creating highly detailed and contextually accurate images from textual prompts. A notable area of clinical application is the representation of surgeon demographics in various specialties, particularly in the context of microsurgery and plastic surgery-related subspecialties.
This cross-sectional study, conducted in June 2024, utilized the latest version of the Copilot Creative Mode powered by DALL-E 3 to generate images of surgeons across various plastic surgery subspecialties. Real-world demographic data from the US, Japan, and Zambia were compared with AI-generated images for an accurate representation analysis.
Five hundred images (350 from various subspecialties and 150 from geographical sources) were analyzed. The AI model predominantly generated images of male and female surgeons with a statistical underrepresentation of female and Black microsurgeons. Geographical prompts influenced the representation, with an overrepresentation of female (64.0 %; p < 0.001) and Black (16.0 %; p < 0.001) plastic surgeons in the US and exclusively Asian surgeons in Japan. Discrepancies were also observed in the depiction of surgical equipment, with the majority of AI-generated microsurgeons inaccurately portrayed using either surgical loupes (46.0 %) or optical microscopes (32.0 %), not with surgical microscopes (4.0 %).
This study revealed significant disparities between AI-generated images and actual demographics in the fields of microsurgery and plastic surgery-related subspecialties, highlighting the need for more diverse and accurate training datasets for AI models.
自2022年11月OpenAI发布ChatGPT以来，生成式人工智能（AI）模型在包括外科手术在内的各个领域引起了广泛关注。这些进展尤其显著地体现在能够根据文本提示创建高度详细且上下文准确的图像。临床应用的一个显著领域是在各个专科中呈现外科医生的人口统计学特征，特别是在显微外科和整形外科相关亚专科的背景下。
这项横断面研究于2024年6月进行，利用由DALL-E 3驱动的最新版本的Copilot创意模式生成各个整形外科亚专科的外科医生图像。将来自美国、日本和赞比亚的真实世界人口统计学数据与人工智能生成的图像进行比较以进行准确的代表性分析。
共分析了500张图像（350张来自各个亚专科，150张来自地理来源）。人工智能模型主要生成的是男性和女性外科医生的图像，女性和黑人显微外科医生的统计代表性不足。地理提示影响了代表性，在美国，女性（64.0%；p<0.001）和黑人（16.0%；p<0.001）整形外科医生的代表性过高，而在日本则完全是亚洲外科医生。在手术设备的描绘上也存在差异，大多数人工智能生成的显微外科医生被不准确地描绘为使用手术放大镜（46.0%）或光学显微镜（32.0%），而不是手术显微镜（4.0%）。
本研究揭示了在显微外科和整形外科相关亚专科领域中，人工智能生成的图像与实际人口统计学之间存在显著差异，突出了为人工智能模型提供更多样化和准确的训练数据集的必要性。
