中国激光, 2020, 47 (10): 1007002, 网络出版: 2020-10-09
基于深度学习的低信噪比下的快速超分辨荧光显微成像 下载: 1323次
Fast Super-Resolution Fluorescence Microscopy Imaging with Low Signal-to-Noise Ratio Based on Deep Learning
摘要
超分辨荧光成像实验的分辨率和成像质量与实验过程中收集到的荧光分子光子数和背景噪声有着密切的关系。为了实现低光子数、高背景光下的快速超分辨荧光显微成像,利用所提卷积神经网络算法实现了对极低信噪比信号的恢复,并结合重构网络进行了超分辨成像。结果表明:利用该方法可以实现荧光信号在低信噪比下的有效恢复,峰值信噪比可达27 dB,明显优于同类的其他两种算法。该方法还可以配合Deep-STORM重构网络在低信噪比下实现快速的超分辨成像。重构结果的归一化均方误差为7.5%,分辨率相较其他算法有明显提升。实验条件下的重构结果验证了该方法的能力,为弱信号下的荧光快速超分辨成像提供了可行方案。
Abstract
The resolution and imaging quality of super-resolution fluorescence imaging significantly depend on the number of fluorescent molecular photons collected during the experiment, as well as the background noise. To obtain fast super-resolution fluorescence microscopy imaging under low photon count and high background light conditions, the proposed convolutional neural network is employed to restore the signal with extremely low signal-to-noise ratio (SNR) and combined with the reconstruction network to perform super-resolution imaging. The results show that the fluorescence signal can be effectively recovered under the condition of low signal-to-noise ratio, the peak signal-to-noise ratio can reach 27 dB, which is significantly better than the other two algorithms. The proposed method can also cooperate with Deep-STORM reconstruction network to obtain fast super-resolution imaging under low SNR conditions. The normalized mean square error of the reconstructed result is 7.5%, and the resolution is significantly improved compared to the other similar algorithms. Additionally, the reconstruction results under experimental conditions verify the ability of the proposed method and provide a feasible solution for fast super-resolution fluorescence imaging under weak signals.
肖康, 田立君, 王中阳. 基于深度学习的低信噪比下的快速超分辨荧光显微成像[J]. 中国激光, 2020, 47(10): 1007002. Xiao Kang, Tian Lijun, Wang Zhongyang. Fast Super-Resolution Fluorescence Microscopy Imaging with Low Signal-to-Noise Ratio Based on Deep Learning[J]. Chinese Journal of Lasers, 2020, 47(10): 1007002.