近年来,荧光显微成像技术由于良好的特异性、高的对比度和信噪比等性能优势,被广泛应用于生物物理学、神经科学、细胞学、分子生物学等生命科学研究的各个领域。然而,传统的荧光显微镜仍然存在分辨率、成像速度、成像视场、光毒性和光漂白等的相互限制,使其在亚细胞结构观测、活体生物超精密成像和分子结构研究领域的应用受到了极大阻碍。由于传统荧光显微镜的局限性,研究人员将目光投向了由数据驱动的深度学习方法。基于深度学习的显微镜的出现,丰富了现有的光学显微成像技术,大数据量的训练突破了传统光学显微镜所能够达到的功能和性能的疆界。本文聚焦基于深度学习的荧光显微成像技术,首先对深度学习的基本原理以及发展过程进行简要概述,随后针对深度学习在荧光显微成像领域近年来的国内外最新成果进行总结,之后通过与传统显微成像系统进行对比,阐述了深度学习在解决荧光显微成像问题上的优越性,最后对深度学习在显微成像技术上的应用前景进行了展望。

In recent years, fluorescence microscopy has been commonly applied in various fields of scientific research, such as biophysics, neuroscience, cell biology, and molecular biology, owing to its specificity, high contrast, and high signal-to-noise ratio. However, traditional fluorescence microscopes have limitations regarding spatial resolution, imaging speed, field of view, phototoxicity, and photobleaching; these limitations compromise their applications in subcellular observation, in vivo imaging, and molecular structure profiling. To moderate such limitations, researchers have adopted data-driven deep learning methods, which can enrich the existing fluorescence microscopy technologies and boost the performance boundary of traditional fluorescence microscopy. This article focuses on the technologies and applications of deep learning based fluorescence microscopy. First, we briefly summarize the basic principle and development path of deep learning technologies; then, we introduce the latest domestic and global progress of deep learning based fluorescence microscopy. Compared with the traditional microscopic imaging system, we show the superiority of deep learning in solving fluorescence microscopy problems. Finally, the future potential of developing deep learning based microscopy is highlighted.