基于人工神经网络对荧光偶极子空间方位的高精度识别

High Precision Identification of Fluorescent Dipole Spatial Orientation Based on Artificial Neural Network

摘要: 单分子定位显微镜具有在纳米尺度上解析生物样品结构细节的能力。然而，传统的方法计算比如高斯模型拟合，Kirchhoff矢量近似和极化法，因复杂度太高或者是需要复杂精密的实验仪器而不利于快速准确的预测定位，限制了单分子定位显微镜的分辨率。最近，人工神经网络技术的发展极大的促进了传统光学显微技术的发展。因此，基于单分子荧光偶极子分子在高NA成像系统下的点分布函数的物理模拟，本文提出并研究了将多层感知器和卷积神经网络两种人工神经网络应用在预测荧光偶极子分子的空间取向。本文的研究显示，相比于多层感知器，卷积神经网络不仅对荧光偶极子空间取向具有更高的准确率，其运行速度和参数量都具有较明显的优势。而且，卷积神经网络在模拟的高噪声预测中有更强的鲁棒性。本文的研究为卷积神经网络在快速准确的单分子荧光成像中应用奠定了基础。

Abstract: Single-molecule localization microscopy has the capability to resolve structural details of biological samples at the nanoscale. However, traditional computational methods such as Gaussian model fitting, Kirchhoff vector approximation, and polarization methods are limited in their ability to predict localization rapidly and accurately, either due to their high complexity or the requirement for complex and precise experimental instruments, thereby constraining the resolution of single-molecule localization microscopy. Recently, the significant development of artificial neural network technology has greatly promoted the advancement of traditional optical microscopy techniques. Therefore, based on the physical simulation of the point spread function of single- molecule fluorophores in high NA imaging systems, this study proposes and investigates the application of two artificial neural networks, namely multilayer perceptrons and convolutional neural networks, in predicting the spatial orientation of fluorescent dipole molecules. The research demonstrates that, compared to multilayer perceptrons, convolutional neural networks not only exhibit higher accuracy in predicting the spatial orientation of fluorescent dipoles but also have distinct advantages in terms of operational speed and parameter count. Moreover, convolutional neural networks show stronger robustness in simulated high-noise predictions. This study lays the groundwork for the application of convolutional neural networks in rapid and accurate single-molecule fluorescence imaging.