Scholarlink

<p>The second near-infrared (NIR-II) fluorescence imaging is a new biomedical imaging method. Compared with the NIR-IIa window (1000–1300 nm), the NIR-IIb window (1500–1700 nm) produces clearer images and better imaging effects. Due to technical limitations, no NIR-IIb molecular probes are used in clinic. In order to get the NIR-IIb images, we translate the given NIR-IIa images into the NIR-IIb images through artificial intelligence. We propose a generative model to complete the above translation process, which is based on the encoder–decoder structure. The decoder is a pre-trained StyleGAN trained only in the NIR-IIb domain. NIR-II images are usually unpaired. In order to achieve unsupervised image translation, we separate the latent code of StyleGAN into two parts: style–content and style–style. The encoder encodes the images into the latent space, and by learning the content representation and style representation in the latent space, the generated images can inherit the content of the NIR-IIa images and the style of the NIR-IIb images. Experiment results show that the proposed model can generate high quality NIR-IIb fluorescence images compared with multiple existing methods.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Learning content and style representation for NIR-II fluorescence image translation

  • 期刊论文

  • 发表日期：

    2025年4月2日

查看全文
  • Xiaoming Yu,
  • Xiaojing Shi,
  • Jie Tian,
  • Zhenhua Hu

摘要

The second near-infrared (NIR-II) fluorescence imaging is a new biomedical imaging method. Compared with the NIR-IIa window (1000–1300 nm), the NIR-IIb window (1500–1700 nm) produces clearer images and better imaging effects. Due to technical limitations, no NIR-IIb molecular probes are used in clinic. In order to get the NIR-IIb images, we translate the given NIR-IIa images into the NIR-IIb images through artificial intelligence. We propose a generative model to complete the above translation process, which is based on the encoder–decoder structure. The decoder is a pre-trained StyleGAN trained only in the NIR-IIb domain. NIR-II images are usually unpaired. In order to achieve unsupervised image translation, we separate the latent code of StyleGAN into two parts: style–content and style–style. The encoder encodes the images into the latent space, and by learning the content representation and style representation in the latent space, the generated images can inherit the content of the NIR-IIa images and the style of the NIR-IIb images. Experiment results show that the proposed model can generate high quality NIR-IIb fluorescence images compared with multiple existing methods.