本文为美国斯坦福大学（作者：Noah Y. Toyonaga）的毕业论文，共41页。

自适应光学技术有可能将深层组织的光学成像分辨率提高到衍射极限。我们研究了神经网络作为一种工具的可能，通过从一个PSF1预测波前相位误差来帮助提高自适应光学系统的收敛速度。自适应光学系统在天文成像中成功地校正了大气条件引起的像差。在这个装置中，人工导航星和波前传感器可用于重建信号波前的空间相位变化。在深层组织中，由于组织的高散射特性，波前传感器无法精确测量相位误差。因此，优化变形镜阵列位置的替代策略是必要的。

以往的研究表明，随机排列反射镜自由度的随机梯度下降法可以使反射镜的峰值强度提高2-5倍。然而，这个过程非常耗时（需要600到3000个摄像机集成才能找到最佳阵列位置）。我们希望训练一个神经网络，直接从失真的PSF图像预测相位失真。本文通过预测相应PSF图像的Zernike和Kolmogorov相位屏，给出了一个概念证明。进一步证明了卷积神经网络结构在相位预测方面优于全连接神经网络结构。最后，我们展示了一个大规模数据收集系统，用于训练基于真实生物组织的神经网络。这项研究代表了首次将神经网络用于显微镜自适应光学的研究，以及首次将卷积神经网络用于任何类型的自适应光学的研究。

Adaptive optics have the potential toimprove the resolution of optical imaging in deep tissue up to the diffractionlimit. We study neural networks as a possible tool to help improve theconvergence rate of adaptive optics systems by predicting the wavefront phaseerror from an abberated PSF1 . Adaptive optics systems have proved successfulin correcting aberrations produced by atmospheric conditions in astronomy imaging.In this setting artificial guide stars and wavefront sensors which can be usedto reconstruct spatial phase variation in the signal wavefront. In deep tissue,given a comparable guide star “particle”, wavefront sensors fail to accuratelymeasure phase error due to the highly scattering nature of the tissue. Thus,alternative strategies to optimize the array positions of the deformable mirrorare necessary. It has previously been shown that a stochastic gradient descentmethod which randomly permutes the degrees of freedom of the mirror canoptimize to an improvement of 2-5x peak intensity. However this process is timeconsuming (necessitating between 600 and 3000 camera integrations to find anoptimal array position). We would like to train a neural network to predictphase distortions directly from distorted PSF images. In this thesis we show aproof of concept by predicting Zernike and Kolmogorov phase screens fromcorresponding PSF images. We further show that convolutional neural networkarchitectures appear superior to fully-connected neural network architecturesin phase prediction. Finally, we demonstrate a system for the mass collectionof data to train a neural network based on real biological tissue. This studyrepresents the first research conducted to use neural networks for adaptiveoptics for microscopy, and the first research to use convolutional neuralnetworks for any type of adaptive optics.

1. 引言

2. 项目背景：神经网络

3. 训练数据的仿真

4. 结构与训练

5. 结果

6. 海量数据采集程序

7. 结论

附录A 测试数据总结

更多精彩文章请关注公众号：