文献速递技术与安全

AI驱动的超写实:电影渲染的下一步?

AI-powered Hyperrealism: Next Step in Cinematic Rendering?
PubMed 文献 · Radiology · 2024-01-01

本研究为探索性研究,于2023年7月27日至8月3日探索Midjourney 5.2(文本到图像生成AI工具)在生成医学图像中的原理、参数与提示工程技术。选取30幅图像(包括眼、皮肤和血管动脉瘤图像),展示该技术生成逼真医学图像的能力,并指出解剖保真度并非总能保持,且在手部手指或精确文本等实体生成方面存在不足。

要点

  • Midjourney 5.2可生成眼、皮肤、血管动脉瘤等逼真医学图像,但解剖保真度并非总能保持。
  • 调整提示词措辞与权重可定制输出图像特征,但生成手部手指或精确文本存在不足。
  • 训练数据过度依赖艺术与摄影,可能导致解剖结果不准确,需放射科医生与AI开发者持续合作。

核心结论

  • AI图像生成通过增加视觉细节和吸引力,在某些应用中有潜力改善三维医学图像,但由于训练数据过度依赖艺术和摄影,可能导致解剖结果不准确,需要放射科医生与AI开发者持续合作,且伦理讨论与版权规定的演变值得密切关注。

研究设计与样本

  • 探索性研究,于2023年7月27日至8月3日探索Midjourney 5.2(文本到图像生成AI工具)的原理、参数和提示工程技术,并给出分步使用说明;选取30幅图像,包括眼、皮肤和血管动脉瘤图像。

局限性

  • Midjourney的模型训练细节保密,据估计其训练数据至少来自网络上数亿张图像。
  • 由于训练数据集不一定基于准确的医学图像,解剖保真度并非总能保持。
  • 该新兴技术在生成手部手指或精确文本等实体方面存在不足。
  • 训练数据过度依赖艺术和摄影,可能导致解剖结果不准确。
  • 伦理讨论与版权规定的演变值得密切关注。

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)。

英文原文(摘要)· 点击展开

Background Recent advancements in artificial intelligence (AI)-powered image generation present opportunities to enhance three-dimensional medical images. Diffusion, an iterative denoising process, represents the standard of many of the current tools used for this purpose. Purpose To demonstrate the current capabilities of diffusion technology by using Midjourney, version 5.2, a text-to-image generative AI tool, and present a practical guide for its use. Materials and Methods This exploratory study investigates the principles, parameters, and prompt engineering techniques for generating images focusing on Midjourney from July 27 to August 3, 2023. Step-by-step instructions show the innate capability of this technology in creating realistic medical images. Results Thirty images were selected, including eye, skin, and vascular aneurysm images. Varying prompt phrasing and weighting techniques allowed for the customization of output image characteristics. Although the details of Midjourney's model training are confidential, it is estimated that it was trained on at least hundreds of millions of images from the web. Anatomic fidelity was not always maintained because the training data set is not necessarily based on accurate medical images. There are shortcomings in this nascent technology regarding its ability to create entities such as digits of the hand or precise text. Conclusion AI image generation has the potential to improve three-dimensional medical images for certain applications through added visual detail and appeal but ongoing collaboration is needed between radiologists and AI developers due to the overreliance on art and photography in the training data, which may result in inaccurate anatomic results. Moreover, the evolving landscape of ethical discussions and copyright stipulations warrants close attention. © RSNA, 2024 Supplemental material is available for this article.

来源:Radiology,查看原文。版权归原作者所有。

引用

  1. Javan R, Mostaghni N. AI-powered Hyperrealism: Next Step in Cinematic Rendering? Radiology. 2024. PMID: 38289206. doi:10.1148/radiol.231971

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。