超分辨荧光显微镜中的解卷积技术及应用(特邀)

Deconvolution in Super-Resolution Fluorescence Microscopy (Invited)

超分辨荧光显微镜突破了光学衍射极限造成的空间分辨率限制,使得生物学家能够在生命体和细胞具有活性的状态下,对其功能与结构进行高精度动态记录,有望揭示更多重要的生命现象细节。然而,由于超分辨荧光显微技术的成像视场、深度、分辨率、速度等不易兼得,所以解卷积作为一种最有效且直接的求解逆问题的框架,被广泛应用于增强超分辨显微镜的时空分辨率。研究人员聚焦于通过相应算法设计实现高质量显微图像的重建,在一定程度上克服了超分辨荧光显微镜的硬件限制,可以更好地恢复生物信息。本文首先介绍了解卷积方法的基本原理及其发展历程,接着列举了不同解卷积技术在不同模态下的重建原理和效果以及这些技术在生物学上的应用,最后总结了基于深度学习的解卷积方法在超分辨荧光显微镜技术上的最新进展和未来的发展潜力,并对包括傅里叶环相关的定量评估图像重建质量的方法的最新进展进行了阐述。

Significance Owing to its non-invasiveness and high specificity,fluorescence microscopy is widely utilized in biomedical research to investigate the structures and functions of biological systems.Limited by the diffraction of light,the resolution of conventional fluorescence microscopy is~250 nanometer(nm)and~800 nm on the lateral and axial axes,respectively,and it cannot resolve nanostructures beyond this limit.To overcome the resolution limit,many super-resolution fluorescence microscopy techniques have been developed,enabling biologists to record the dynamics of the fine structures of organisms and cells in their active states.This offers the potential to elucidate the crucial details of biological phenomena.Nevertheless,in super-resolution fluorescence microscopy,trade-offs exist between resolution,speed,and imaging depth.Although these trade-offs can be moderated by optimizing the microscopy hardware,certain strict physical limitations cannot be easily overcome.Therefore,enhancing microscopy performance via computational imaging methods is particularly important.For instance,the application of deconvolution algorithms can transcend physical limits without changing the optical hardware,thereby improving the dissection of biological information.Progress This review introduces the technical principles of various deconvolution methods.Deconvolution techniques are applied to four modes of super-resolution fluorescence microscopy:structured illumination microscopy(SIM),image scanning microscopy(ISM),stimulated emission depletion(STED)microscopy,and super-resolution optical fluctuation imaging(SOFI).Various modalities have been used for live cell imaging applications.For example,researchers have designed deconvolution algorithms to eliminate the reconstruction artifacts produced during the reconstruction of SIM and to improve its resolution.Additionally,for SOFI,deconvolution techniques can be applied as pre-or post-processing steps to further enhance the efficiency of utilizing statistical information and to improve resolution.The recently developed advanced deconvolution algorithm,sparse deconvolution,is stable and robust to various noise conditions and can effectively improve the three-dimensional resolution two-fold.Furthermore,it can be combined with different variants of fluorescence microscopy to enhance their contrast and resolution in situ without any changes.Owing to significant advances in the corresponding super-resolution reconstruction techniques,live-cell super-resolution microscopy has been effectively enhanced.In the outlook section,considering the unrolling algorithm as an example,this review discusses the prospects of deconvolution methods based on deep learning. The combination of deep learning algorithms and microscopy imaging techniques may become a future development trend in the field of live-cell super-resolution microscopy. This review briefly describes the Fourier ring correlation (FRC) image resolution measurement method and its application in image reconstruction. Finally, a rolling FRC (rFRC) method is introduced to quantitatively detect the reconstruction uncertainties of super-resolution techniques at the corresponding super-resolution scale.Conclusions and Prospects Owing to hardware limitations, extensive super-resolution microscopy methods have introduced computational steps to achieve the optimal quality of super-resolution imaging. This review can serve as a bridge between the super-resolution microscopy and computation communities to facilitate the application of novel computational techniques toward improved resolution, accuracy, and image processing.